Daily Papers

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded "prompt templates", i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, i.e. imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, a few lines of DSPy allow GPT-3.5 and llama2-13b-chat to self-bootstrap pipelines that outperform standard few-shot prompting (generally by over 25% and 65%, respectively) and pipelines with expert-created demonstrations (by up to 5-46% and 16-40%, respectively). On top of that, DSPy programs compiled to open and relatively small LMs like 770M-parameter T5 and llama2-13b-chat are competitive with approaches that rely on expert-written prompt chains for proprietary GPT-3.5. DSPy is available at https://github.com/stanfordnlp/dspy

Speculative decoding (SD) addresses the high inference costs of LLMs by having lightweight drafters generate candidates for large verifiers to validate in parallel. Existing draft-verify methods use binary decisions: accept or fully recompute. Yet we find that many rejected tokens can be verified correctly by a slim submodel derived from the full verifier via intra-model routing, instead of the full verifier. This motivates our slim-verifier to handle tokens requiring moderate verification resources, reducing expensive large-model calls. We propose Verification via Intra-Model Routing for Speculative Decoding (VIA-SD), a multi-tier framework using a routed slim-verifier. Draft tokens are processed hierarchically: direct acceptance for high-confidence cases, slim-verifier regeneration for medium-confidence cases, and full-model verification for uncertain cases. Across four representative tasks and multiple model families, VIA-SD reduces rejection rates by 0.10-0.22 and delivers 10-20% speedups over strong SD baselines, while achieving 2.5-3x acceleration over non-drafting decoding. Moreover, VIA-SD is compatible with existing SD frameworks without modifying their training procedures. Our results suggest multi-tier SD as a general paradigm for scalable and efficient LLM inference. Project page: https://zju-xyc.github.io/VIA-SD-Project-Page/

Probability-only models, which TypeSafe calls System One models, return calibrated probabilities for fixed choices in milliseconds and generate no text. We study one such model, Jev, through two tasks that require decisions under tight constraints. In bullet chess, a bot that places Jev's judgment inside Stockfish search alongside an opening book and endgame tablebases climbs above a 2200 Lichess bullet rating against other bots. Live model calls are too slow for search, so we distill pairwise judgments into a compact evaluator that runs at every position. We then ask how best to spend a fixed labeling budget when an LLM, Qwen3-32B, is available as a second teacher. In chess, averaging both judges' labels beats spending the whole budget on Qwen alone by 9.6 Elo (95% interval 4.3 to 14.9), and the gain replicates on fresh openings; a second answer from the same judge is no substitute, and Jev is the strongest partner for Qwen among the models tested. In passage reranking, Jev's labels alone train a reranker that scores as high as Qwen's, from 21 minutes of API calls instead of 5.1 GPU-hours, and adding Qwen gains at most a few thousandths in ranking quality. Search supplies the lookahead, distillation makes the judgment cheap enough to use at every position, and an LLM partner pays off in chess.

Generative AI is being leveraged to solve a variety of computer-use tasks involving desktop applications. State-of-the-art systems have focused solely on improving accuracy on leading benchmarks. However, these systems are practically unusable due to extremely high end-to-end latency (e.g., tens of minutes) for tasks that typically take humans just a few minutes to complete. To understand the cause behind this and to guide future developments of computer agents, we conduct the first study on the temporal performance of computer-use agents on OSWorld, the flagship benchmark in computer-use AI. We find that large model calls for planning and reflection account for the majority of the overall latency, and as an agent uses more steps to complete a task, each successive step can take 3x longer than steps at the beginning of a task. We then construct OSWorld-Human, a manually annotated version of the original OSWorld dataset that contains a human-determined trajectory for each task. We evaluate 16 agents on their efficiency using OSWorld-Human and found that even the highest-scoring agents on OSWorld take 1.4-2.7x more steps than necessary.

Many disciplines pose natural-language research questions over large document collections whose answers typically require structured evidence, traditionally obtained by manually designing an annotation schema and exhaustively labeling the corpus, a slow and error-prone process. We introduce ScheMatiQ, which leverages calls to a backbone LLM to take a question and a corpus to produce a schema and a grounded database, with a web interface that lets steer and revise the extraction. In collaboration with domain experts, we show that ScheMatiQ yields outputs that support real-world analysis in law and computational biology. We release ScheMatiQ as open source with a public web interface, and invite experts across disciplines to use it with their own data. All resources, including the website, source code, and demonstration video, are available at: www.ScheMatiQ-ai.com

We show that GPT-4's reasoning and planning capabilities extend to the 1993 first-person shooter Doom. This large language model (LLM) is able to run and play the game with only a few instructions, plus a textual description--generated by the model itself from screenshots--about the state of the game being observed. We find that GPT-4 can play the game to a passable degree: it is able to manipulate doors, combat enemies, and perform pathing. More complex prompting strategies involving multiple model calls provide better results. While further work is required to enable the LLM to play the game as well as its classical, reinforcement learning-based counterparts, we note that GPT-4 required no training, leaning instead on its own reasoning and observational capabilities. We hope our work pushes the boundaries on intelligent, LLM-based agents in video games. We conclude by discussing the ethical implications of our work.

Multi-turn LLM agents interleave model calls with external tool invocations, shifting serving from stateless request processing to stateful program execution. Serving these workloads requires scheduling, KV-cache management, and routing policies that use program-level context, including turn dependencies, tool-induced gaps, and reusable KV state. Evaluating such policies directly on real systems is costly, since each design point may require dedicated accelerator time across arrival rates, model scales, serving-instance counts, and memory hierarchies. Simulation offers a scalable alternative, but existing LLM serving simulators target stateless request-level workloads and therefore omit the core dynamics of agent serving: multi-turn program execution, cross-turn cache locality, and KV-cache residency during tool gaps. We present AGENTSERVESIM, a hardware-aware simulator for multi-turn LLM agent serving. AGENTSERVESIM evaluates serving policies at program granularity through composable modules: a Program Orchestrator preserves program identity and turn order, a Tool Simulator materializes tool-induced gaps, a Session-Aware Router maintains program-to-instance affinity for cache-aware dispatch, and a KV Residency Model tracks policy-defined KV placement across HBM, host DRAM/CXL, and eviction. Across real serving deployments and hardware configurations, AGENTSERVESIM reproduces real-system behavior within 6% error across key performance metrics while running entirely on commodity CPUs. These results show that AGENTSERVESIM enables controlled, repeatable exploration of agent-serving policies without requiring exhaustive deployment on costly accelerators.

Deciding whether to call a tool is a core competence of an LLM agent, and a costly one to get wrong: needless calls add latency, accrue cost, and may trigger irreversible side effects, while missing calls leave the model confidently wrong on questions it could only answer through tool-calls. Models manage this balance poorly, both over-using and under-using tools. Existing methods such as post-training and prompt engineering are expensive and difficult to modify at inference time. We show that whether an instruction-tuned model calls a tool can be controlled by a single linear direction in its residual stream, extracted without any training from the model's own tool-use preference signal and turned into an inference-time intervention with no prompt change. Adding the direction with strength α moves the call rate monotonically from near 0% to over 90% while keeping calls well-formed. The steering works in both directions: dialing it down suppresses calls, and dialing it up induces new calls that land precisely on the questions the model cannot answer from its own knowledge. We also show that the direction generalizes to unseen tools with strength comparable to each tool's own direction and without favoring any specific tool choice. With live tool execution, a single sweep of the steering traces a cost/accuracy Pareto frontier and nearly doubles open-domain QA accuracy (0.29 ! rightarrow ! 0.56); the same recipe transfers across a diverse range of models spanning dense, MoE, and multimodal architectures, without any training. Our code is publicly available at https://github.com/YuqiChen4188/Steering-Tool-Use-Propensity.

Giving an agent a file about a named expert can supply hard-to-find material, produce a recognizable persona, or change what the agent decides. These are different claims. We test each one. mimeo is an open-source tool that finds a person's public work, checks each extracted quotation against the cached source text, and writes a file an agent can load. Eight logged builds averaged 38 model calls; the check rejects 13.2% of extracted quotations. We tested four expert files with one coding-agent harness. Knowledge access was clearest: mimeo answered all 20 obscure, quotation-heavy questions; no closed-book condition answered more than 10. Keyword search (BM25) over the same pages answered 15-17, a gap this sample cannot resolve. Grounding showed one clear benefit: personas written from model memory misstated a documented position on 1-4 of 20 answers under every grader; the plain agent and mimeo never did. Every persona was easy to spot on short open prompts, and adding task material lowered identification by 18-23 points. mimeo was no more identifiable than a from-memory profile. Judgment transfer remained unresolved because both tests hit their ceiling: every condition found 94-97% of the problems planted in engineering tasks and scored 94-100% on 16 new application scenarios. An AI-judged "sounds like the expert" score changed with the judge: two of four preferred answers based on a model's stereotype, while two found no difference on the same text. That is a caution against relying on a single AI judge. The evidence supports mimeo as a compact, inspectable reference on a person, not as a demonstrated transfer of their judgment. Toolkit and expert profiles: https://github.com/K-Dense-AI/mimeo

LLM agents increasingly take on long-running tasks whose history grows far beyond a single model context window. Existing approaches compress earlier interactions or extract selected information into fixed memory representations, committing to what to preserve before future needs are known. We present Scroll, a context manager that treats each agent session as an executable Session Environment. The environment is backed by an append-only Event Log and a sandboxed, persistent Python kernel. The kernel maintains a typed namespace across model calls, allowing tool outputs, retrieved history, and derived state to be bound to variables rather than serialized into the prompt at each call. Model-written code searches, materializes, and transforms session state through exec; only explicitly printed projections enter the model's working view for the next call. Context management thus becomes a programming task that inherits the improving coding abilities of LLMs, while the Event Log preserves lossless historical ground truth. As the working view approaches its budget, stale spans are evicted but remain recoverable: an eviction index keeps compact landmarks tied to exact Event Log addresses, so that the agent navigates directly to evicted regions instead of searching the full log. With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points.

GPT-6 Astra exhibits a remarkable ability to generate numerical robot actions, extending its role beyond high-level planning. To assess Astra's capabilities as general-purpose embodied policies, we conduct comprehensive evaluations across six domains, examining direct control, cooperation with learned policies, and feedback-driven adaptation. In gripper manipulation, Astra can correct task targets and prepare contact conditions for subsequent policy execution; hybrid control with π0.5 achieves 48% success on the evaluated RoboDojo subset. In dexterous manipulation, hybrid control achieves 50% success in ten experience-guided DexJoCo trials, while direct in-hand control struggles to coordinate finger contacts. In mobile manipulation, hybrid control reaches 38.7% success on the evaluated RoboCasa365. In navigation, Astra leads our local comparisons, reaching 92% success on RxR instruction following and 82% on HM3D object search, although search incurs substantial detours. In locomotion, dense motion-reference generation remains unreliable: none of five sequential attempts on a single obstacle course reaches the goal, despite improvements in stability and forward progress. In humanoid loco-manipulation, Astra exceeds baseline methods on 13 of 30 HumanoidBench tasks with pretrained whole-body controllers. These findings reveal a gap between useful task decisions and reliable physical control. Inference latency further constrains practical control: across 50 RoboDojo instances per condition, policy-assisted and direct control consume 624.8 million and 1.132 billion tokens. A 30-second locomotion run requires 250 model calls averaging 39.86 seconds each, with physics paused during inference.

Visual agents for symbolic graphics reasoning depend on executable world representations, yet continued improvement remains bottlenecked by ambiguous supervision from final renders. Recent agentic systems refine scenes through repeated interaction, but they lack mechanisms to recursively extract actionable evidence from execution traces, overlooking which proposed symbolic edits survived, disappeared, or interfered during program execution. To address this, we propose VETR (Verified Execution Trajectories for Refinement), a framework that enables agentic recursive self-improvement over executable world representations for symbolic graphics reasoning. Within each task, VETR retains aligned program/render records with bounded runtime observations and introduces an execution refiner that compiles observed writes and state changes into constrained alternative programs. The refiner validates their endpoints and preservation conditions against captured runtime state; the ordinary visual verifier then assesses goal relevance. Repair construction makes no additional language-model calls, but can require additional executions. Evaluations on BlenderGym (280 tasks) and BlenderBench (27 released tasks) with GPT-5.6-luna and Qwen3.8-27B show that VETR consistently outperforms existing refinement strategies, including BlenderAlchemy and VIGA. Controlled ablations evaluate the contributions of execution refinement and role-asymmetric evidence organization under fixed generator-interaction caps.

Large language model (LLM) agents are evolving from request-response assistants into long-running software actors: they maintain state across model calls, fork subtasks, wait for external events, request human authority, generate tools, and perform side effects that must be resumed and audited. This paper presents Agent libOS, a library-OS-inspired runtime substrate for LLM agents. Agent libOS runs above a conventional host operating system; it does not implement hardware drivers, kernel-mode isolation, or a POSIX-compatible operating system. Instead, it treats an agent as an AgentProcess: a schedulable execution subject with process identity, parent-child lineage, lifecycle state, a tool table derived from an AgentImage, typed Object Memory, explicit capabilities, human queues, checkpoints, events, and audit records. Its central design rule is tools are libc-like wrappers; runtime primitives are the authority boundary. Filesystem access, object access, sleeps, human approval, JIT tool registration, and external side effects are checked at primitive boundaries under explicit capabilities and policy. We describe the design, threat model, Python prototype, and safety-oriented evaluation. The current prototype implements async scheduling, namespace-local Object Memory, runtime-integrated human approval, one-shot permission grants, per-process working directories, shell and image-registration primitives, Deno/TypeScript JIT tools over a libOS syscall broker, filesystem/object bridge tools, an injectable Resource Provider Substrate, deterministic demos, real-model smoke scripts, and 123 regression tests at the time of writing. Rather than improving planner accuracy, Agent libOS demonstrates a runtime substrate in which long-running LLM agents can be scheduled, authorized, resumed, and audited without treating tool dispatch as the trust boundary.

Spatial intelligence is becoming a foundation for embodied agents, robotic planning, and multimodal assistants. To improve the spatial reasoning ability of VLM agents, existing work has mainly followed two lines. One line uses post-training methods, such as supervised fine-tuning and reinforcement learning. Another line adopts an agentic paradigm in which the model calls external spatial tools, such as depth estimation and 3D reconstruction tools, to gather intermediate spatial evidence. We study a complementary and underexplored route: Can a frozen VLM agent improve its spatial reasoning through parameter-update-free self-evolution, without depending on external expert spatial tools at inference time? We present Spatial Memory Agent (SMA), an experience-grounded runtime framework that converts verified spatial experience into reusable transferable lessons. In a verifiable spatial environment, SMA queries the frozen VLM, obtains a predicted answer and reward, and uses verifier-guided reflection to distill compact transferable lessons from spatial experience. SMA further assigns each lesson a Transfer Reliability Score (TRS), which is initialized uniformly and calibrated from later retrieval outcomes as visit evidence of future transfer reliability. During read-only deployment, SMA retrieves lessons by semantic filter and similarity-TRS combined ranking, allowing the retrieved memory to guide frozen model inference. Across five representative spatial benchmarks and four base VLMs, SMA achieves the highest macro average in every base-model block and the best accuracy among the evaluated methods in most of the 20 evaluations, establishing a practical parameter-update-free path for spatial self-evolution across the evaluated frozen model scales and environments.

Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost-quality trade-off: escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete. Each switch requires the receiver to continue a non-native trajectory produced by another model. We study how this handoff affects quality and cost, and how varying the trajectory information inherited by the receiver changes the outcome. Using pairs of low-cost, low-capability (LC) and high-cost, high-capability (HC) models from the Claude and GPT families, we vary handoff direction, timing, and interface, comparing full-trajectory transfer, compaction, and trajectory removal while preserving the repository state. Across both model families, full-trajectory escalation recovers less than half of the LC-to-HC quality gap while incurring a substantial cost premium. We term this cost-quality penalty the handoff tax. By contrast, downshift offers a favorable cost-quality point. Interestingly, the preferred interface also reverses with direction: reducing LC-model trajectory information improves escalation quality, whereas removing the HC-model trajectory reduces downshift quality.

Autoregressive decoding (AR) with Transformer models is memory bandwidth bound at single stream inference, the typical deployment regime for on device text to speech (TTS). Real time streaming with Qwen3-TTS requires more than 200 sequential model calls per second, dominated by the inner loop MultiCodeDecoder that emits the 15 residual vector quantization (RVQ) codes per 80 ms audio frame. We propose RVQ position aware speculative decoding for the MultiCodeDecoder, attaining 2.47 accepted tokens per model call at 5times10^{-4} percent added parameters and 10 to 20 percent per round speculation/verification overhead, reducing real time synthesis from 200 to 88 sequential model calls per second. The scheme is distributionally lossless under the deployed top-k sampling, and WER parity with the original system is consistent with this guarantee. We deliver 2 to 2.2x speedup for RVQ token generation with Qwen3-TTS 0.6B on recent iPhone and Apple Silicon Mac devices.

A language-model agent is jointly defined by its model and its harness, the executable program that organizes model calls, tool use, and information flow. Because different tasks call for different ways of organizing these operations, the harness needs to be adapted using feedback from the task at hand. We introduce harness learning, which trains a proposer model to revise a solver's harness using execution feedback. We formulate this process as meta-learning over executable programs, with harness revisions playing the role of weight updates in gradient-based adaptation. We train the proposer with reinforcement learning, using the task performance of revised harnesses as the reward. At test time, the proposer uses feedback from successive executions on a new task to refine the harness, without performing any parameter-space update. Experiments on reasoning and multi-hop question answering show that harness learning improves revision quality and that the ability to adapt at test time transfers to unseen tasks. Policies trained on individual revisions can continue improving harnesses over multiple rounds, while the benefits of training on revision sequences vary across settings. These findings suggest a path towards continually learning agents that turn accumulated experience into generalizable improvements.

FSampler is a training free, sampler agnostic execution layer that accelerates diffusion sampling by reducing the number of function evaluations (NFE). FSampler maintains a short history of denoising signals (epsilon) from recent real model calls and extrapolates the next epsilon using finite difference predictors at second order, third order, or fourth order, falling back to lower order when history is insufficient. On selected steps the predicted epsilon substitutes the model call while keeping each sampler's update rule unchanged. Predicted epsilons are validated for finiteness and magnitude; a learning stabilizer rescales predictions on skipped steps to correct drift, and an optional gradient estimation stabilizer compensates local curvature. Protected windows, periodic anchors, and a cap on consecutive skips bound deviation over the trajectory. Operating at the sampler level, FSampler integrates with Euler/DDIM, DPM++ 2M/2S, LMS/AB2, and RES family exponential multistep methods and drops into standard workflows. FLUX.1 dev, Qwen Image, and Wan 2.2, FSampler reduces time by 8 to 22% and model calls by 15 to 25% at high fidelity (Structural Similarity Index (SSIM) 0.95 to 0.99), without altering sampler formulas. With an aggressive adaptive gate, reductions can reach 45 to 50% fewer model calls at lower fidelity (SSIM 0.73 to 0.74).

Although LLM agents can leverage tools for complex tasks, they still need memory to maintain cross-turn consistency and accumulate reusable information in long-horizon interactions. However, retrieval-based external memory systems incur low online overhead but suffer from unstable accuracy due to limited query construction and candidate filtering. In contrast, many systems use repeated large-model calls for online memory operations, improving accuracy but accumulating latency over long interactions. We propose LightMem, a lightweight memory system for better agent memory driven by Small Language Models (SLMs). LightMem modularizes memory retrieval, writing, and long-term consolidation, and separates online processing from offline consolidation to enable efficient memory invocation under bounded compute. We organize memory into short-term memory (STM) for immediate conversational context, mid-term memory (MTM) for reusable interaction summaries, and long-term memory (LTM) for consolidated knowledge, and uses user identifiers to support independent retrieval and incremental maintenance in multi-user settings. Online, LightMem operates under a fixed retrieval budget and selects memories via a two-stage procedure: vector-based coarse retrieval followed by semantic consistency re-ranking. Offline, it abstracts reusable interaction evidence and incrementally integrates it into LTM. Experiments show consistent gains across model scales, with an average F1 improvement of about 2.5 over A-MEM on LoCoMo, while achieving higher efficiency and low median latency (83 ms for retrieval and 581 ms end-to-end).

Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.

Power sampling improves large language model (LLM) reasoning at test time by sampling from a sharpened version of the base model's distribution. Its sampler, PowerMH, is slow because its Metropolis--Hastings (MH) updates are sequential: each accept/reject decision sets the state from which the next proposal is generated. Prior work reduces this cost by changing the sampler or the model (e.g., sequential Monte Carlo, approximate lookahead, distillation). We propose Predictive SubTree Prefetching (PreSTO), which accelerates PowerMH by prefetching future proposals and provably leaves its MH transition kernel unchanged. PreSTO views future MH outcomes as a tree of accept/reject decisions: in each model call, it generates the proposals of a prefetched subtree of this tree in one batch, then follows the realized path with the original MH acceptance rule. We also observe that MH reaches only a sparse, nearly unary part of each prefetched subtree, we further design a predictive traversal rule that chooses the subtree to maximize the predicted number of completed transitions. Across seven datasets and five base LLMs, PreSTO achieves median speedups of 1.71times, 1.91times, and 2.33times on PowerMH and its follow-ups EntropyCut and MultiTryMH, respectively; on the first two, it makes a median of 3.9times fewer model calls, each only 2.1times longer. On LCB V6 with Qwen3.5-9B, the predictive traversal rule is a further 1.25times--1.37times faster than the default breadth-first traversal.

An AI agent run is no longer a function call. It is a cascade of prompts, model calls, tool invocations, file and network effects and sub-agents. When such a run misbehaves, or an auditor asks what happened, logs and traces answer poorly: they are mutable, partial and cannot be re-executed. This technical report describes NovaFabric, an open-source (Apache-2.0) black-box recorder for AI agents that turns a run into a sealed, replayable, diffable and auditable evidence capsule. The design rests on two formats only. The Run Capsule is a directory of flat files and the source of truth; every database is a rebuildable index over capsules. The Evidence Bundle is the signed export for an auditor. Everything else follows existing standards: OpenTelemetry GenAI, OpenLineage, W3C PROV, in-toto/DSSE, RFC 3161 and CycloneDX. A single-node run is the smallest case of a distributed one; the schema supports a capsule hierarchy for Slurm arrays and Kubernetes jobs, though automatic parent finalization is library-only today. The report covers the architecture and the engines behind the verb chain capture, seal, replay, diff and audit: fail-closed mocked replay; a diff gate with a stable exit-code contract; DSSE signatures that bind their intent and a capsule root that commits to file names; a capsule treated as untrusted input; signer identity-trust levels with offline TSA and CRL checking; maker-checker promotion against an operator-held trust list; and a deny-by-default authorization table for every dashboard route. It presents no new experiments; the evaluation is in arXiv:2609.12582. Every count except the private ADR tally is generated by a probe script run on the public source at the v0.107.0 tag (150 top-level commands, 84 Python subpackages, 16,721 test functions), every capability carries an explicit maturity label, and honest-limits boxes state where the system stops.

Scaling inference-time compute for Large Language Models (LLMs) has unlocked unprecedented reasoning capabilities. However, existing inference-time scaling methods typically rely on inefficient and suboptimal discrete search algorithms or trial-and-error prompting to improve the online policy. In this paper, we propose nabla-Reasoner, an iterative generation framework that integrates differentiable optimization over token logits into the decoding loop to refine the policy on the fly. Our core component, Differentiable Textual Optimization (DTO), leverages gradient signals from both the LLM's likelihood and a reward model to refine textual representations. nabla-Reasoner further incorporates rejection sampling and acceleration design to robustify and speed up decoding. Theoretically, we show that performing inference-time gradient descent in the sample space to maximize reward is dual to aligning an LLM policy via KL-regularized reinforcement learning. Empirically, nabla-Reasoner achieves over 20% accuracy improvement on a challenging mathematical reasoning benchmark, while reducing number of model calls by approximately 10-40% compared to strong baselines. Overall, our work introduces a paradigm shift from zeroth-order search to first-order optimization at test time, offering a cost-effective path to amplify LLM reasoning.

We introduce self-jailbreaking, a threat model in which an aligned LLM guides its own compromise. Unlike most jailbreak techniques, which often rely on handcrafted prompts or separate attacker models, self-jailbreaking requires no external red-team LLM: the target model's own internal knowledge suffices. We operationalize this via Self-Jailbreaking via Lexical Insertion Prompting (\textsc{SLIP)}, a black-box algorithm that casts jailbreaking as breadth-first tree search over multi-turn dialogues, incrementally inserting missing content words from the attack goal into benign prompts using the target model as its own guide. Evaluations on AdvBench and HarmBench show SLIP achieves 90--100\% Attack Success Rate (ASR) (avg.\ 94.7\%) across most of the eleven tested models (including GPT-5.1, Claude-Sonnet-4.5, Gemini-2.5-Pro, and DeepSeek-V3), with only {sim}7.9 LLM calls on average, 3--6times fewer than prior methods. We evaluate existing defenses, show that regex-based approaches are evaded by prompt paraphrasing, and propose the Semantic Drift Monitor (SDM) defense that tracks SLIP's embedding-space trajectory, achieving 76\% detection at 5\% FPR. However, SDM remains insufficient against adaptive attack strategies, underscoring the need for more advanced defense mechanisms tailored to the self-jailbreaking threat surface. We release our code for reproducibility.

AI agents have been developed for complex real-world tasks from coding to customer service. But AI agent evaluations suffer from many challenges that undermine our understanding of how well agents really work. We introduce the Holistic Agent Leaderboard (HAL) to address these challenges. We make three main contributions. First, we provide a standardized evaluation harness that orchestrates parallel evaluations across hundreds of VMs, reducing evaluation time from weeks to hours while eliminating common implementation bugs. Second, we conduct three-dimensional analysis spanning models, scaffolds, and benchmarks. We validate the harness by conducting 21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000. Our analysis reveals surprising insights, such as higher reasoning effort reducing accuracy in the majority of runs. Third, we use LLM-aided log inspection to uncover previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks. We share all agent logs, comprising 2.5B tokens of language model calls, to incentivize further research into agent behavior. By standardizing how the field evaluates agents and addressing common pitfalls in agent evaluation, we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world.

Large Language Models (LLMs) have achieved remarkable success in natural language understanding and generation, but their deployment is constrained by high computational demands. Deploying smaller LLMs directly on the edge can circumvent this, but with degraded accuracy. Deploying smaller cloud-based big LLMs preserves performance, but at the cost of expensive per-token computation. We present a distributed inference framework, , that integrates speculative decoding (SD) across edge and cloud. A compact draft model deployed on the edge generates candidate tokens rapidly, and a large verifier model on the cloud validates these tokens in parallel. Accepted tokens are retained, while only rejections trigger verifier correction, substantially reducing the number of cloud queries. Our plug-and-play design shifts the bulk of computation to the edge, significantly lowers inference time and cloud cost, and preserves the accuracy of the big model without any retraining requirement. Our approach demonstrates a practical path toward scalable, cost-efficient, and accurate deployment of LLMs in real-world environments. Experimental results across multiple Natural Language Processing tasks using SpecBench and CNN/Dailymail datasets demonstrate that reduces the cloud model calls by 76% with zero loss in accuracy as compared to the full model.

LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and τ-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.

Deep research requires connecting multiple steps of search, reasoning, and computation, with intermediate findings guiding what to investigate next. As investigations grow, agents must maintain research state across model calls and check findings before errors propagate into subsequent work. We introduce PANINI-DR, a multi-agent harness that separates coordination of the overall investigation from the research and verification of individual subquestions. A shared workspace, built on the Generative Semantic Workspace representation, records subquestions, their dependencies, findings, and supporting evidence. Researchers investigate individual subquestions, while a separate verification agent checks proposed answers against those subquestions and their evidence. Verification feedback guides further investigation, and recorded findings inform subsequent assignments. We evaluate PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven state-of-the-art deep-research harnesses, all using a shared language-model backbone and frozen corpora. PANINI-DR achieves the strongest average performance across the three benchmarks. On BrowseComp-Plus, it achieves 73.3% accuracy, outperforming the next-best harness by 16.0 points. Its advantage is strongest on hard questions, where it answers nearly twice as many as the next-best harness. Ablations further show that structured planning and persistent state improve performance, as does checking each finding against its cited evidence before it is reused. Because a model can answer correctly from memory alone, we also analyze the evidence retained by each harness. PANINI-DR achieves the highest answer recoverability from saved evidence on all three benchmarks, while complementary claim-support audits show stronger evidence-backed answers. Accounting for this evidence changes how harnesses compare and further strengthens PANINI-DR's advantage.

As LLM capabilities advance, agents are tackling increasingly complex tasks over longer horizons. Their growing interaction histories make memory compaction essential for staying within context windows and reducing prefill cost. Existing methods summarize the history or compress its KV cache, often adding model computation to preserve information for future requests. A new user request can change which history matters, but reassessing that history with the model requires re-encoding it if the KV cache has expired. Past attention provides signals of historical importance and dependencies between messages, while relevance to the current task must be assessed using the new user request. We introduce ReCAP, a memory compaction method that stores attention-derived importance scores and dependency links in a lightweight, persistent context graph. For each new request, ReCAP combines stored importance with relevance cues from the request and follows dependency links to select messages and their supporting context, without additional model calls for selection. Compared with Codex's default summarization-based compaction, ReCAP reduces estimated latency for compaction and cold restoration by approximately 95% on both Qwen3-Coder and gpt-oss. It also roughly halves the historical context per call on SWE-Together at comparable task quality and improves accuracy on the code tasks of Lost-in-Conversation over full history by 19.8 and 41.2 points.

Vision-language models (VLMs) achieve remarkable performance on 2D medical images, such as X-rays or pathology images, while 3D VLMs that directly process CT/MRI volumes lag behind due to limited 3D training data and the predominance of 2D foundation architectures. To leverage powerful 2D VLMs for analyzing medical volumes, which contain hundreds of slices, it is therefore necessary to focus only on those few slices, which are diagnostically relevant. To enable such a prioritization, we propose XGLASS, a method that transfers medical expertise from 2D X-rays to 3D CT for task-aware slice prioritization. XGLASS uses disease-conditioned importance maps derived from language-aligned radiologist gaze on X-rays and spatially aligns them with coronal CT projections, bringing gaze priors and CT content into a shared 2D representation. The fused gaze-prior/CT projection is then used to predict a disease-specific ordering of axial slices, prioritizing diagnostically relevant slices. XGLASS significantly improves budgeted retrieval over 2D and 3D-native baselines, achieving a top-10 hit rate up to 75.9% for the best-performing disease. In practice, this means that after inspecting only the top 10 slices (4.17% of a 240-slice scan), XGLASS identifies a diagnostically relevant slice in more than 75% of cases. For downstream slice-wise finding classification, an XGLASS-selected 30-slice candidate set reduces model calls by 87.5% while improving F1 by 11.3 percentage points over all-slice inference, 4.2 points over the strongest 2D selector, and remains competitive with 3D-native selectors while requiring up to 130.8 times less time and 10.9 times less GPU memory per finding, supporting its use as a lightweight preselector for volumetric CT analysis. The code and pretrained checkpoint are available on https://github.com/michaelgloeckler/XGLASS.

Long-horizon LLM agents act on an append-only interaction history that is re-sent to the model at every step, so the context and its cost grow with steps until the sessions exceed the context window. Existing methods manage the context through context requirement prediction, relying on additional model calls, heuristic rules, or trained policies. However, these predictive approaches introduce runtime overhead, invalidate prefix caches, and permanently discard content with no guarantee of recovery. To overcome these limitations, we introduce ReFold: a training-free rendering layer that preserves the underlying interaction history while compressing only the model's rendered context. It removes two kinds of inter-turn redundancy without an auxiliary predictor: content an earlier turn already displayed, replaced by a stub, and turns the agent itself reports finished, folded into a one-line note. Both operators use chunked rendering, rewriting the cached prefix once every few steps rather than at every step. Every removal is strictly reversible, a wrong removal costs one restore from the history rather than permanent content loss. Because it operates at the rendering layer, ReFold is plug-and-play across standard ReAct-style harnesses. Evaluations across five long-horizon benchmarks and two frontier LLMs demonstrate that ReFold reduces token consumption by up to 2.5x and halves the KV-cache memory per session without degrading task success rates. Under capped context budgets, it avoids up to 92% of forced compactions. Under concurrent serving workloads, it reduces request queuing delays by up to 100%, accelerating inference by up to 1.7x, while cutting inference costs by up to 3.4x.

Locating relevant evidence in long inputs is a core capability of Multimodal Large Language Models. Prior studies attribute this capability to retrieval heads, a sparse set of attention heads that transfer information from the context to the output. However, these analyses are confined to a limited set of modalities and the corresponding models. Moreover, how retrieval heads can be leveraged to enhance the model calls for further exploration. In this paper, we ask whether this mechanism expands together with the input modalities of the model. Taking audio-visual input as a newly added modality, we construct paired inputs in which the same answer is either spoken or shown on screen. With the evidence content fixed, retrieval heads remain sparse in the new modality and are shared to a large extent across modalities. Across 18 language, vision-language, and omni-modal checkpoints, omni-modal models largely preserve the retrieval heads of their backbones. Ablating these heads eliminates nearly all audio-visual retrieval. The heads shared across modalities make up fewer than 2\% of all heads, yet ablating them also degrades document understanding, spatial perception, and visual mathematics. Building on these findings, we turn retrieval heads into a target of optimization by training their attention with a ranking objective. This raises the page Recall@1 of Qwen3-VL-8B on MMDocIR by 6.03 points and improves document retrieval across three model families. Functional attention heads can thus serve both as a lens on multimodal models and as a target for improving them.

High-quality articulated assets are essential for embodied perception, simulation, and manipulation, yet constructing them at scale remains difficult. Manual annotation and procedural modeling require substantial human effort. Recent agentic systems automate asset construction through program synthesis, but repeat costly reasoning and refinement for every new asset. Our key observation is that assets within a category can share construction logic despite variations in structure and appearance. Agentic synthesis and procedural generation are therefore complementary: agents can construct reusable programs, while procedural execution reuses the resulting logic across many samples. We therefore present RFW-A, which combines the two in a new articulated-asset generation paradigm. Starting from existing articulated programs, RFW-A organizes reusable components and their dependencies into category-level generators. Agents test and refine these generators across sampled configurations and motion states. Once constructed and validated, the generators produce diverse assets through procedural execution with no further model calls or token charges. The estimated API cost of generator authoring averages 10.62, amortizing to 0.05 per asset over 200 assets, versus $1.24 for the sampled per-asset agentic synthesis baseline. Using RFW-A, we built RFW-A-400K, containing over 400K sim-ready articulated assets across more than 500 categories. Experiments show high structural and kinematic validity and stable execution across physics simulators. Beyond asset validity, training with RFW-A-400K improves part-level articulation estimation on seen and unseen categories, while policies trained in simulation with retrieved and program-edited assets transfer zero-shot to real articulated objects.

Test-time scaling has emerged as a promising approach for improving code generation by exploring large solution spaces at inference time. However, existing methods often rely on public test cases that are unavailable in practice, or require extensive LLM inference for candidate selection, leading to significant token consumption and time overhead. We present DiffCodeGen, a novel test-time scaling method for code generation based on coverage-guided differential analysis. DiffCodeGen generates diverse code candidates using various sampling and prompting strategies, then applies coverage-guided fuzzing to synthesize inputs without requiring any existing tests or large language models. By executing all candidates on these inputs, DiffCodeGen captures their dynamic behavior and clusters candidates based on behavioral similarity. DiffCodeGen selects the medoid of the largest cluster as the final output. Unlike prior test-time scaling methods that invoke additional LLM inference for candidate selection, DiffCodeGen performs selection without any extra model calls, incurring little to no additional token consumption. DiffCodeGen is fully asynchronous, naturally suited to the current trend of agentic coding, and is thus efficient and highly scalable. We evaluate DiffCodeGen across 4 large language models, demonstrating consistent improvements over baselines. Compared to state-of-the-art test-time scaling methods, DiffCodeGen achieves competitive or superior performance while using only a fraction of time and tokens. DiffCodeGen is model-agnostic and can be combined with reasoning models to further boost performance.

Agent harnesses have substantially improved performance on long-horizon tasks by coordinating agent interactions with the environment. However, reinforcement learning through complex harnesses remains largely unexplored, as scaling such training to long-horizon agent tasks introduces fundamental challenges. In this work, we present a unified black-box RL framework for stable and scalable optimization of general agents through complex harnesses. Concretely, we first build a sandbox-based execution infrastructure that isolates task environments and harnesses within temporary sandboxes for large-scale concurrent rollouts. We then decouple policy optimization from opaque harness execution and place a serving proxy at the model boundary to capture model calls. To reconstruct multi-turn trajectories and improve training efficiency, we organize the captured calls into prefix trees and further adapt both critic-based PPO and critic-free GRPO to optimize over the recovered tree structure. Meanwhile, we maintain training-inference consistency throughout the optimization process. Finally, we introduce mix-harness training, allowing a single model to be jointly optimized by heterogeneous harnesses. With Qwen3-30A3B, black-box RL improves Pass@1 on ClawGym-Bench by 9.98 and 14.81 points through OpenClaw and Claude Code, respectively, while remaining stable over 200-400 optimization steps. Moreover, the framework yields consistent gains on more challenging tasks such as JobBench and OfficeQA. Overall, our framework enables effective, stable, and scalable optimization of general agents through black-box harnesses, supporting unified training across heterogeneous execution systems.

Classification accuracy and true-positive and false-positive rates alone do not determine whether restarting an AI agent improves success. Using an exact alarm-to-action identity for an independent clean restart with zero monitoring overhead, we show analytically that a monitor with fixed 90% accuracy can swing from a five-percentage-point loss to a 35-point gain in success rate, depending solely on the replacement attempt's residual-budget success probability. We introduce two techniques for evaluating monitors and progress signals in decision terms. First, a decision-sufficiency criterion, a binary-action specialization of Blackwell's comparison of experiments, identifies which progress representations or context compressions preserve the information needed to choose optimally between the specified actions, independent of trace reconstruction fidelity. Second, a finite-sample, model-free decision audit uses simultaneous binomial intervals over existing agent traces, without additional model calls, to certify a preference for "restart" or "continue" under stated sampling assumptions, or return "abstain" when evidence is insufficient. In a study of 194 externally evaluated model executions on 19 selected SymPy issues from SWE-bench Verified, with a retrospective audit of 130 of these traces, we find a 2.29-point apparent restart gain traced to a single surviving failed attempt, a 42.29-point loss on a second model, no positive replication of the initial gain on prospective task cohorts, and zero of 103 live comparisons meeting the specified certification bar. These results motivate evaluating agent monitors through recovery opportunity, decision value, and statistical support, rather than classification performance alone.

Multiple-choice QA benchmarks usually evaluate small language models (SLMs) as direct answerers, but deployed language-model systems increasingly rely on external scaffolds such as tools, code, and repeated model calls. We introduce Code-Guided Reasoning (CGR), an evaluation protocol and generated-program resource for measuring when executable reasoning scaffolds improve SLM performance on MCQA tasks. CGR standardizes six components: a normalized item interface, a direct solver prompt, a generator prompt, a Python scaffold, solver-call and extraction helpers, and a three-channel result record. On 20,498 retained result rows from a locally prepared MCQA bundle and six metadata-registered solver models, the observed non-zero-baseline partition shows 66.21% macro assisted accuracy versus 38.11% direct accuracy, a +28.10 percentage-point difference with a pair-bootstrap interval of [20.32, 36.43]. Under a stricter Ab > 30% direct-signal gate, the macro difference is +14.11 points. These estimates are descriptive. Assisted inference uses a larger solver-call budget, answer extraction is brittle, Time-MQA contains the observed regressions, and some generated programs violate the no-hard-coding instruction. CGR provides the trace package needed to interpret these results, including direct, assisted, and generator-side answers, partition definitions, generated programs, response metadata, and audits.

Retrosynthesis, which aims to find a route to synthesize a target molecule from commercially available starting materials, is a critical task in drug discovery and materials design. Recently, the combination of ML-based single-step reaction predictors with multi-step planners has led to promising results. However, the single-step predictors are mostly trained offline to optimize the single-step accuracy, without considering complete routes. Here, we leverage reinforcement learning (RL) to improve the single-step predictor, by using a tree-shaped MDP to optimize complete routes. Specifically, we propose a novel online training algorithm, called Planning with Dual Value Networks (PDVN), which alternates between the planning phase and updating phase. In PDVN, we construct two separate value networks to predict the synthesizability and cost of molecules, respectively. To maintain the single-step accuracy, we design a two-branch network structure for the single-step predictor. On the widely-used USPTO dataset, our PDVN algorithm improves the search success rate of existing multi-step planners (e.g., increasing the success rate from 85.79% to 98.95% for Retro*, and reducing the number of model calls by half while solving 99.47% molecules for RetroGraph). Additionally, PDVN helps find shorter synthesis routes (e.g., reducing the average route length from 5.76 to 4.83 for Retro*, and from 5.63 to 4.78 for RetroGraph).

Mathematical word problem-solving has long been recognized as a complex task for small language models (SLMs). A recent study hypothesized that the smallest model size, needed to achieve over 80% accuracy on the GSM8K benchmark, is 34 billion parameters. To reach this level of performance with smaller models, researcher often train SLMs to generate Python code or use tools to help avoid calculation errors. Additionally, they employ ensembling, where outputs of up to 100 model runs are combined to arrive at a more accurate result. Result selection is done using consensus, majority vote or a separate a verifier model used in conjunction with the SLM. Ensembling provides a substantial boost in accuracy but at a significant cost increase with multiple calls to the model (e.g., Phi-GSM uses top-48 to boost the performance from 68.2 to 81.5). In this work, we present Orca-Math, a 7-billion-parameter SLM based on the Mistral-7B, which achieves 86.81% on GSM8k without the need for multiple model calls or the use of verifiers, code execution or any other external tools. Our approach has the following key elements: (1) A high quality synthetic dataset of 200K math problems created using a multi-agent setup where agents collaborate to create the data, (2) An iterative learning techniques that enables the SLM to practice solving problems, receive feedback on its solutions and learn from preference pairs incorporating the SLM solutions and the feedback. When trained with Supervised Fine-Tuning alone, Orca-Math achieves 81.50% on GSM8k pass@1 metric. With iterative preference learning, Orca-Math achieves 86.81% pass@1. Orca-Math surpasses the performance of significantly larger models such as LLAMA-2-70B, WizardMath-70B, Gemini-Pro, ChatGPT-3.5. It also significantly outperforms other smaller models while using much smaller data (hundreds of thousands vs. millions of problems).

Inference-time scaling can enhance the reasoning capabilities of large language models (LLMs) on complex problems that benefit from step-by-step problem solving. Although lengthening generated scratchpads has proven effective for mathematical tasks, the broader impact of this approach on other tasks remains less clear. In this work, we investigate the benefits and limitations of scaling methods across nine state-of-the-art models and eight challenging tasks, including math and STEM reasoning, calendar planning, NP-hard problems, navigation, and spatial reasoning. We compare conventional models (e.g., GPT-4o) with models fine-tuned for inference-time scaling (e.g., o1) through evaluation protocols that involve repeated model calls, either independently or sequentially with feedback. These evaluations approximate lower and upper performance bounds and potential for future performance improvements for each model, whether through enhanced training or multi-model inference systems. Our extensive empirical analysis reveals that the advantages of inference-time scaling vary across tasks and diminish as problem complexity increases. In addition, simply using more tokens does not necessarily translate to higher accuracy in these challenging regimes. Results from multiple independent runs with conventional models using perfect verifiers show that, for some tasks, these models can achieve performance close to the average performance of today's most advanced reasoning models. However, for other tasks, a significant performance gap remains, even in very high scaling regimes. Encouragingly, all models demonstrate significant gains when inference is further scaled with perfect verifiers or strong feedback, suggesting ample potential for future improvements.

The development of GUI agents could revolutionize the next generation of human-computer interaction. Motivated by this vision, we present MAI-UI, a family of foundation GUI agents spanning the full spectrum of sizes, including 2B, 8B, 32B, and 235B-A22B variants. We identify four key challenges to realistic deployment: the lack of native agent-user interaction, the limits of UI-only operation, the absence of a practical deployment architecture, and brittleness in dynamic environments. MAI-UI addresses these issues with a unified methodology: a self-evolving data pipeline that expands the navigation data to include user interaction and MCP tool calls, a native device-cloud collaboration system routes execution by task state, and an online RL framework with advanced optimizations to scale parallel environments and context length. MAI-UI establishes new state-of-the-art across GUI grounding and mobile navigation. On grounding benchmarks, it reaches 73.5% on ScreenSpot-Pro, 91.3% on MMBench GUI L2, 70.9% on OSWorld-G, and 49.2% on UI-Vision, surpassing Gemini-3-Pro and Seed1.8 on ScreenSpot-Pro. On mobile GUI navigation, it sets a new SOTA of 76.7% on AndroidWorld, surpassing UI-Tars-2, Gemini-2.5-Pro and Seed1.8. On MobileWorld, MAI-UI obtains 41.7% success rate, significantly outperforming end-to-end GUI models and competitive with Gemini-3-Pro based agentic frameworks. Our online RL experiments show significant gains from scaling parallel environments from 32 to 512 (+5.2 points) and increasing environment step budget from 15 to 50 (+4.3 points). Finally, the native device-cloud collaboration system improves on-device performance by 33%, reduces cloud model calls by over 40%, and preserves user privacy.

Large language models (LLMs) exhibit latent multi-token prediction (MTP) capabilities despite being trained solely for next-token generation. We propose a simple, training-free MTP approach that probes an LLM using on-the-fly mask tokens drawn from its embedding space, enabling parallel prediction of future tokens without modifying model weights or relying on auxiliary draft models. Our method constructs a speculative token tree by sampling top-K candidates from mask-token logits and applies a lightweight pruning strategy to retain high-probability continuations. During decoding, candidate predictions are verified in parallel, resulting in lossless generation while substantially reducing the number of model calls and improving token throughput. Across benchmarks, our probing-based MTP consistently outperforms existing training-free baselines, increasing acceptance length by approximately 12\% on LLaMA3 and 8--12\% on Qwen3, and achieving throughput gains of up to 15--19\%. Finally, we provide theoretical insights and empirical evidence showing that decoder layers naturally align mask-token representations with next-token states, enabling accurate multi-step prediction without retraining or auxiliary models.

Agent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision, none of which behaves like a deep-learning optimizer for the skill, and none of which reliably improves over its starting point under feedback. We argue the skill should instead be trained as the external state of a frozen agent, with the same discipline that makes weight-space optimization reproducible. SkillOpt is, to our knowledge, the first systematic controllable text-space optimizer for agent skills: a separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document, and an edit is accepted only when it strictly improves a held-out validation score. A textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta update make skill training stable while adding zero inference-time model calls at deployment. Across six benchmarks, seven target models, and three execution harnesses (direct chat, Codex, Claude Code), SkillOpt is best or tied on all 52 evaluated (model, benchmark, harness) cells and beats every per-cell competitor among human, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. On GPT-5.5 it lifts the average no-skill accuracy by +23.5 points in direct chat, by +24.8 inside the Codex agentic loop, and by +19.1 inside Claude Code. Transfer experiments further show that optimized skill artifacts retain value when moved across model scales, between Codex and Claude Code execution environments, and to a nearby math benchmark without further optimization.

LLM agents increasingly rely on iterative execution to solve tasks through planning, tool use, state updates, and agent collaboration. While this design enables flexible automation, it also creates a new class of failures: an agent may repeatedly execute model calls, tools, workflow transitions, or agent handoffs when the feedback path is not effectively bounded. We call this problem Infinite Agentic Loops (IALs). IALs are not ordinary programming loops; they arise from the interaction between agent logic, framework semantics, runtime observations, and termination mechanisms. Such failures can amplify a single request into long running model and tool execution, causing cost exhaustion, model denial of service, context growth, and repeated external side effects. We propose IAL-Scan, a static analysis tool for detecting IAL failures in real-world LLM agent projects. IAL-Scan abstracts heterogeneous agent code into a framework independent Agent IR, builds an Agentic Loop Dependence Graph (ALDG) to recover explicit and framework induced feedback paths, and checks whether these paths can repeatedly reach costly or state growing operations without an effective bound. We evaluate IAL-Scan on 6,549 LLM agent repositories. It reports 74 potential findings, among which manual review confirms 68 IAL failures across 47 projects, achieving 91.9% precision.

Embodied agents are typically built as hand-designed compositions of perception, memory, planning, and action modules. This modularity exposes a large architectural design space, but current systems still rely on researcher intuition to choose where information is stored, how observations are processed, and how model calls are connected. Agent Architecture Search (AAS) automates such design for text-domain agents, but has not been systematically evaluated on perceptual embodied agents through simulator rollouts. We study this transfer. We introduce AgentCanvas, a typed-graph runtime that hosts embodied executors as editable node-and-wire programs with simulator-aware execution and episode-level logs, and KDLoop, a coding-agent search procedure that cycles through proposal, critique, experiment, and distillation, with triggered reflection after stalls. We evaluate three AAS variants across four embodied executors spanning vision-language navigation, embodied question answering, and language-conditioned manipulation. The resulting 3x4 matrix shows that architecture-level search can produce deployable and directional success-rate gains on embodied tasks, while one apparent high-scoring candidate is rejected as leak-bearing. At the same time, the experiments expose constraints that are muted in text-domain AAS: optimization signals can be masked by rollout noise, search can become trapped in local edit basins, and episode-level credit assignment only partially emerges even when detailed logs are available. These results characterize both the promise and the current limits of automated architecture search for embodied agents.

Budget-constrained agentic search arises when an LLM agent must refine candidates under a small evaluation budget, because validation is expensive, generation requires multiple model calls, or both. In this regime, standard MCTS allocates budget poorly: exploration bonuses dominate at low visit counts, unpromising siblings are expanded before promising chains can deepen, and branching is independent of node quality. We introduce ExTS, a tree-search policy that treats expansion itself as a value-of-information decision. ExTS combines three mechanisms: discriminative reward shaping to separate candidates under narrow score distributions, a stochastic virtual child that estimates the value of creating a new branch from the parent's reward history, and quality-conditioned branching that expands only when a node's score justifies the budget cost. Across prompt optimization, code generation, molecular structure elucidation, and agentic workflow optimization, ExTS is competitive with or improves over task-specific tree-search baselines, with an average relative gain of +5.5% using a single fixed configuration. We further introduce pilot-run diagnostics that characterize what makes budget-constrained agentic search problems structurally different from one another, providing both understanding of the problem space and practical guidance for adaptation.

LLM agents that cache recovery suggestions from API errors can skip re-derivation in later episodes, spending fewer tokens and fewer model calls on constraints they have already learned. Server-side data drift turns those cached fixes into silent failures, and the usual remedy, re-deriving on every episode, gives the savings back. We introduce invalidation contracts, a protocol layer that attaches version stamps and cacheability hints to every recovery suggestion so the client can evict stale entries without trial and error, and keep the rest. The contract decomposes realized savings into two independent factors: validity, the fraction of cached suggestions that remain correct after a drift event, and compliance, the fraction the planner applies on the first attempt. Validity depends only on the protocol and is vendor-independent. Compliance depends on the planner model: identical wire bytes yield 100% first-try compliance on Claude Haiku 4.5 and 11% or below on Claude Sonnet 5, which exhibits input-schema conservatism, refusing fixes that add fields the original request did not contain. We evaluate across seven models, three serving paths, two domains, and approximately 9,400 episodes. Row-level invalidation raises compliance by 0 to 66.7 percentage points across the seven models, 55.6 to 66.7 on three, and recovers 29-33% of baseline token cost on four of seven models, while table-level invalidation destroys co-located entries and drops post-drift first-try rates to 0% on five of seven. Eviction precision is 1.00 at row granularity on every model under the row-level oracle of Section 4.1. The contract adds 15% to response payload. Version-stamp validity is deterministic by construction and produced identical results across every model and serving path, with zero contract failures in the entire evaluation.

Large language model (LLM)-powered coding agents have made rapid progress in automating software engineering tasks, yet repository-level issue resolution remains challenging. Beyond generating a plausible patch, an agent must localize relevant code across interdependent files and maintain repository context that is both sufficient and focused. Code graphs expose non-local relations, but linear text interfaces obscure their topology; rendering the full repository graph yields visual representations that are too dense to perceive reliably, whereas a one-shot local view becomes stale as exploration proceeds. We present RepoAtlas, a training-free module that maintains evolving multimodal repository views through a select--project--refresh loop over a repository code graph. RepoAtlas combines evidence from the issue with the agent's current exploration state to select a task-relevant region under a fixed budget, projects the selected structure into complementary visual and textual representations, and refreshes the view when changes in the exploration state render it outdated. We evaluate RepoAtlas on SWE-bench Verified, where it improves the resolve rate by 2.4 points while reducing input tokens and model calls by 5.8\% and 7.8\% on average, relative to the strongest multimodal graph baseline, with consistent gains across three models of different families and scales.

Verified solutions are not equally useful for preparing reasoning models for reinforcement learning (RL). We present a comprehensive study of route diversity, the variation in the sequences of reasoning steps in supervised fine-tuning (SFT) data, and propose a lightweight, rule-based fingerprint to select for it. From one pool at one budget, with matched training recipes and checkpoints, selecting diverse rather than similar routes improves post-RL problem coverage across puzzles and mathematics, including on problems harder than those seen in either training stage. In synthetic experiments, route-diverse SFT improves OLMo3-7B's pass@8 by 16.9 points on environments held out from SFT. In a single-model condition, where one model writes every candidate, diverse selection gains up to 6.2 points of mean pass@8 across 10 mathematics benchmarks. Pre-RL diagnostics suggest why: diverse SFT can produce both successful and failed attempts on more prompts despite slightly lower mean accuracy, giving group-relative RL more prompts with a learning signal. On 3 open-source corpora, our CPU-only selector, without model calls, outperforms more expensive alternatives in every comparison of mean post-RL performance. These results identify reasoning-route diversity as a practical criterion for selecting SFT data that better prepares models for RL.

Tool calling, invoking external tools on demand, is central to agentic LLMs, yet the mechanism that decides whether a model calls a tool or responds directly remains poorly understood. Agentic prompts are long and heavily scaffolded, combining role instructions, tool schemas, format templates, and the user's request across hundreds of tokens, creating a noisy, highly entangled context in which no single controllable variable for mechanistic analysis is obvious. To obtain such a variable, we propose a method that converts complex agentic prompts into minimal contrastive pairs in which a single request verb determines the tool-call decision: replacing an execution-verb (e.g., write) with an analysis-verb (e.g., discuss) reliably flips the decision, suggesting it is mediated by a compact internal state. We construct 500 such paired prompts across Python, Java, and C++ (300 for mechanistic analysis, 200 held out for evaluation). We trace the decision to a vector, μ_Δ, that is both causally necessary and sufficient and generalizes beyond the discovery prompts to native multi-turn τ^2-Bench trajectories and verb-free requests. Behavioral ablations show that the scaffold establishes a tool-call prior; Transcoder decomposition then reveals that analysis verbs suppress this prior through features signaling that tool use is unnecessary, whereas execution verbs largely leave it intact. Downstream scaffold-reading attention heads and MLP features read out the resulting state, and the same mechanism recurs across seven models from the Qwen, Mistral, and Granite families. Our code is available at https://github.com/XijieGo/MI4ToolCalling.

Robot learning in simulation depends on the objects the simulator offers. Many tasks need objects with separate parts, joints that allow the required motion, and physical properties that remain valid under contact. Existing methods recover this structure anew for every image: generative models predict parts and joints that mostly fail to settle or move in simulation, and general-purpose agents need a long session of model calls for each photograph. AffordCraft builds such an asset from a single RGB image and a task instruction by retrieval instead of generation: it locates the object and the part to operate, selects a matching entry from a library of articulated assets, and fits it to the image while keeping its parts and joints intact. Without any box or mask marking the object, AffordCraft produces a physically valid asset for 1,703 of 2,000 photographs from 31 categories. Five generative methods pass on at most 45% of the same photographs and, at the median, need 10 to 78 times our GPU time per valid asset. On 50 cluttered images, 162 of 237 annotated objects pass the same physical test after automatic detection. Growing the library from 141 to 11,372 entries needs no change to the method and raises category coverage from 46% to 100% and the share of selections with the requested label from 18% to 51%. We also build manipulation tasks from the constructed assets, both with single objects and in composed scenes; policies trained on scripted demonstrations complete both kinds of tasks from initial states unseen in training.

LLM routing dynamically selects and coordinates heterogeneous language models across diverse tasks. However, existing approaches are largely built around a predefined execution paradigm, such as single-round, multi-round, or agentic routing, which fixes how model calls are organized and limits adaptation to different query needs. We propose RouteWeaver, a unified routing framework that integrates these routing modes within a single LLM policy, jointly learning which mode to use and how to select models within that mode. To jointly optimize mode selection and within-mode execution, we develop COMET-GRPO (COordinated Mode-Execution Training), which decouples the trajectory-level advantage into separate signals for mode selection and within-mode execution, enabling more targeted credit assignment to each decision level. We further introduce progressive exploration, which gradually shifts training from externally assigned modes to autonomous selection, so that the router first learns how to execute each mode before deciding when to use it. We evaluate RouteWeaver across mathematics, coding, knowledge, logical reasoning, and factual recall. RouteWeaver achieves an average accuracy of 61.4%, an 11.03% relative improvement over the strongest baseline, and outperforms all compared baselines in each of the five domains. Ablations show that both COMET-GRPO and progressive exploration are important for adaptive routing; without progressive exploration, the routing policy collapses entirely to single-round execution.

Evaluating task-oriented dialogue agents requires judging not merely whether a reply reads well but whether each turn advances the underlying workflow state correctly--a distinction conventional holistic LLM judges can miss because they evaluate the available context as a single unit and require one or more full-model calls per turn. We propose SAGE (State-Grounded Abstention-Aware Evaluation), which compiles a workflow specification and per-turn state diff into atomic, schema-grounded criteria and routes each through a cascade of symbolic and encoder/NLI verifiers that abstain rather than guess, aggregating criterion verdicts into a turn-level decision with an evidence trace. Its recommended operating point, SAGE-Core, decides 81--91% of criteria with only the compiler, symbolic rules, and on-device encoders--at zero paid LLM cost--while SAGE-LLM adds an optional focused-LLM fallback for open-class criteria. Across four slices spanning MultiWOZ, Schema-Guided Dialogue, and ABCD, no evaluated LLM-as-a-judge baseline--including a state-aware GPT-4.1 judge and cheaper GPT-4.1-mini variants--significantly exceeds SAGE-Core on any slice, even though the GPT-4.1 G-Eval judge costs 4.7--8.0 per 1,000 turns to SAGE-Core's 0. A two-annotator human audit (n=200, κ=0.94) confirms strong label fidelity on the transcript-visible failure classes--where, excluding the weak-salience IUV class, SAGE-Core is statistically tied with the strongest LLM judge--and honestly scopes ignored-user-value as a state-consistency signal with weak broad-human salience. We analyze construct-validity limits from injected failures and partial symbolic circularity.

Autoregressive (AR) language models generate tokens sequentially, so generating long outputs is costly. Speculative decoding mitigates this by using a lightweight drafter to propose multiple tokens for the target AR model to verify in parallel, while a diffusion-based drafter can produce an entire draft block in one forward pass. In models with AR and diffusion modes, one network both drafts and verifies, but drafting still waits for verification, so the standard schedule needs two sequential model calls per iteration. Existing one-call schedules avoid the wait but incur input size quadratic in the speculative width. Motivated by this efficiency bottleneck, we introduce Packed-SS, a self-speculative decoding schedule that verifies and drafts in a single model call with input size linear in the speculative width, while preserving the target model's distribution. The key idea is to let the diffusion drafter run one call ahead: while verifying the current draft, the model simultaneously constructs the next draft from an older verified prefix and keeps its unresolved suffix for the next iteration. Under a mild monotonicity assumption on the drafter, we further show that Packed-SS requires no more model calls than the standard two-call decoding in the greedy setting. Experiments on Nemotron 3B/8B and SetDLM 1.7B show higher throughput and lower mean wall time than both baselines under greedy decoding.

LLM routing matters most in long-horizon applications such as coding agents, deep research systems, and computer-use agents, where a single user request triggers many model calls. Routing each call to the cheapest sufficient model can cut costs without sacrificing quality, yet existing router benchmarks evaluate routers only on one-shot prompts. They never expose the router-visible prefix at an intermediate agent step, never test whether a cheaper replacement preserves downstream task success, and often rely on online LLM judges at evaluation time. We introduce TwinRouterBench, a step-level routing benchmark with two tracks. The static track provides 970 router-visible prefixes from 520 instances across SWE-bench, BFCL, mtRAG, QMSum, and PinchBench, each paired with an execution-verified target tier estimated under a released downgrade-and-cascade protocol; scoring is deterministic arithmetic over tier labels, trajectory membership, and token costs, with no online evaluator-side LLM judge. The dynamic track supplies a harness that runs routers on the full 500-case SWE-bench Verified suite; in this paper we report a 100-case held-out evaluation disjoint from the static SWE supervision split. At each LLM call the router selects a concrete model from a locked pool, and success is measured by official task resolution and realized API spend. The two tracks support fast offline iteration followed by end-to-end validation under live agent execution. Code and data are available at https://github.com/CommonstackAI/TwinRouterBench.

Recent work on activation and latent steering has demonstrated that modifying internal representations can effectively guide large language models (LLMs) toward improved reasoning and efficiency without updating model parameters. However, most existing approaches rely on fixed steering policies and static intervention strengths, which limit their robustness across problem instances and often result in over- or under-steering. We propose Adaptive Test-time Latent Steering (ATLAS), a lightweight framework that dynamically controls steering decisions at inference time using a trained, lightweight verifier over the latent states. Given intermediate hidden states, the verifier predicts the quality of ongoing reasoning and adaptively selects which steering action to apply, enabling per-example and per-step adjustment with minimal overhead. ATLAS provides a unified framework for combining learned latent verification with test-time activation steering, enabling adaptive reasoning control without additional LLM decoding or inference-time process reward model calls. Experiments on multiple mathematical and coding reasoning benchmarks show that ATLAS consistently outperforms both vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage. These results demonstrate that verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality. All source code will be publicly available.

Robotic manipulation often requires remembering interactions that are no longer visible. We introduce Latent Transition Memory (LTM), which stores history as compact latent transitions between consecutive observations. Instantiated on π_{0.5} as LTM-π, our method uses a frozen latent action encoder, a lightweight adapter, and causal KV caching for incremental inference without memory-specific semantic annotations or external reasoning-model calls. With DreamDojo transitions, LTM-π achieves 58.54% success on RoboMME and 80.5% on LIBERO-Mem, performing best among baselines without external reasoning models or memory-specific annotations. Ablations support the contribution of transition content beyond temporal cues and the efficiency of delayed memory admission. Scaling RoboMME demonstrations tenfold improves success from 57.67% to 87.50%, while a Cup Swap diagnostic highlights the role of process supervision in temporal composition. Under a 200-token memory budget, success remains 84.00%, versus 87.50% without compression. Finally, real-robot counting success improves from 11.25% to 82.50%. Together, these results demonstrate the effectiveness, scalability, and efficiency of latent transition memory, highlighting its capacity for complex temporal information, bounded-context operation, and real-world robotic control.

Computer-using agents drive real software through the screen -- clicking and typing -- but they solve every task from scratch: asked to repeat a task, an agent re-reads the screen, re-reasons every tap, and pays the full cost again. We present PreAct, which lets such an agent get faster on tasks it has done before. The first time it succeeds, PreAct compiles the run into a small state-machine program-states that check the screen, transitions that act-and on later runs replays it directly instead of invoking the agent 8.5-13x faster, with no per-step language-model calls. Replay is not blind: at each step PreAct checks that the screen matches what the program expects before acting, and hands control back to the agent the moment something is off. PreAct applies the same discipline when deciding what to keep: a freshly compiled program enters the store only if, re-run from a clean state, an independent evaluator confirms it solved the task-catching programs that replay to their last step yet leave the task undone. Across a mobile, a desktop, and a web benchmark, this store-time check separates repeated runs that improve from ones that degrade as faulty programs accumulate, worth 1.75-2.6 tasks per benchmark, the same direction on all three; a fallback that explores afresh when no program fits brings PreAct level with a strong record-and-replay baseline. We also report what did not matter: prompt wording, runtime guardrails, and whether a language model or a plain embedding retriever selects which program to reuse.

Despite rapid progress in Multimodal Large Language Models (MLLMs), visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable resource for spatial reasoning. We study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

Agentic applications based on large language models increasingly rely on multi-step interaction loops involving planning, action execution, and environment feedback. While such systems are now deployed at scale, improving them post-deployment remains challenging. Agent trajectories are voluminous and non-deterministic, and reviewing each one, whether through human review or auxiliary LLMs, is slow and cost-prohibitive. We propose a lightweight, signal-based framework for triaging agentic interaction trajectories. Our approach computes cheap, broadly applicable signals from live interactions and attaches them as structured attributes for trajectory triage, identifying interactions likely to be informative without affecting online agent behavior. We organize signals into a coarse-grained taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls. In a controlled annotation study on τ-bench, a widely used benchmark for tool-augmented agent evaluation, we show that signal-based sampling achieves an 82\% informativeness rate compared to 74\% for heuristic filtering and 54\% for random sampling, with a 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains, confirming that signals provide genuine per-trajectory informativeness gains rather than merely oversampling obvious failures. These results show that lightweight signals can serve as practical sampling infrastructure for agentic systems, and suggest a path toward preference data construction and post-deployment optimization.

Chaining language model (LM) calls as composable modules is fueling a new powerful way of programming. However, ensuring that LMs adhere to important constraints remains a key challenge, one often addressed with heuristic "prompt engineering". We introduce LM Assertions, a new programming construct for expressing computational constraints that LMs should satisfy. We integrate our constructs into the recent DSPy programming model for LMs, and present new strategies that allow DSPy to compile programs with arbitrary LM Assertions into systems that are more reliable and more accurate. In DSPy, LM Assertions can be integrated at compile time, via automatic prompt optimization, and/or at inference time, via automatic selfrefinement and backtracking. We report on two early case studies for complex question answering (QA), in which the LM program must iteratively retrieve information in multiple hops and synthesize a long-form answer with citations. We find that LM Assertions improve not only compliance with imposed rules and guidelines but also enhance downstream task performance, delivering intrinsic and extrinsic gains up to 35.7% and 13.3%, respectively. Our reference implementation of LM Assertions is integrated into DSPy at https://github.com/stanfordnlp/dspy

Language Model Programs, i.e. sophisticated pipelines of modular language model (LM) calls, are increasingly advancing NLP tasks, but they require crafting prompts that are jointly effective for all modules. We study prompt optimization for LM programs, i.e. how to update these prompts to maximize a downstream metric without access to module-level labels or gradients. To make this tractable, we factorize our problem into optimizing the free-form instructions and few-shot demonstrations of every module and introduce several strategies to craft task-grounded instructions and navigate credit assignment across modules. Our strategies include (i) program- and data-aware techniques for proposing effective instructions, (ii) a stochastic mini-batch evaluation function for learning a surrogate model of our objective, and (iii) a meta-optimization procedure in which we refine how LMs construct proposals over time. Using these insights we develop MIPRO, a novel algorithm for optimizing LM programs. MIPRO outperforms baseline optimizers on five of seven diverse multi-stage LM programs using a best-in-class open-source model (Llama-3-8B), by as high as 13% accuracy. We have released our new optimizers and benchmark in DSPy at http://dspy.ai

A coding agent must emit a valid tool call--a parseable invocation of a tool in the provided schema--before the harness can execute its chosen action. We study how local serving stacks affect this protocol step and show that measured outcomes can depend on the serving layer rather than model behavior alone. In Ollama, the default tools= request is gated per model by a static template flag: some models are accepted and return calls as text, some return native tool_calls, while Phi-3 and Gemma-3 are rejected before inference. In our harness, rejection and retry exhaustion are not preserved as structured failure metadata, so downstream analysis can misclassify them as model non-calls and naively report 0% fidelity. Adding a text tool list while retaining the native channel recovers much of the measured fidelity for accepted models, whereas a uniform text protocol reduces fidelity for Llama-3.2, which has native tool-call support. Cross-stack probes on Ollama, llama.cpp, vLLM, and SGLang show different handling of the same request. Constrained decoding removes parse failures but can induce non-termination, and turn-pooled versus per-instance estimates differ by up to about 55 points. We conclude with a checklist for treating serving behavior as part of the evaluation protocol.

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

The growing number of parameter-efficient adaptations of a base large language model (LLM) calls for studying whether we can reuse such trained adapters to improve performance for new tasks. We study how to best build a library of adapters given multi-task data and devise techniques for both zero-shot and supervised task generalization through routing in such library. We benchmark existing approaches to build this library and introduce model-based clustering, MBC, a method that groups tasks based on the similarity of their adapter parameters, indirectly optimizing for transfer across the multi-task dataset. To re-use the library, we present a novel zero-shot routing mechanism, Arrow, which enables dynamic selection of the most relevant adapters for new inputs without the need for retraining. We experiment with several LLMs, such as Phi-2 and Mistral, on a wide array of held-out tasks, verifying that MBC-based adapters and Arrow routing lead to superior generalization to new tasks. We make steps towards creating modular, adaptable LLMs that can match or outperform traditional joint training.

The deployment of agent systems in an enterprise environment is often hindered by several challenges: common models lack domain-specific process knowledge, leading to disorganized plans, missing key tools, and poor execution stability. To address this, this paper introduces Routine, a multi-step agent planning framework designed with a clear structure, explicit instructions, and seamless parameter passing to guide the agent's execution module in performing multi-step tool-calling tasks with high stability. In evaluations conducted within a real-world enterprise scenario, Routine significantly increases the execution accuracy in model tool calls, increasing the performance of GPT-4o from 41.1% to 96.3%, and Qwen3-14B from 32.6% to 83.3%. We further constructed a Routine-following training dataset and fine-tuned Qwen3-14B, resulting in an accuracy increase to 88.2% on scenario-specific evaluations, indicating improved adherence to execution plans. In addition, we employed Routine-based distillation to create a scenario-specific, multi-step tool-calling dataset. Fine-tuning on this distilled dataset raised the model's accuracy to 95.5%, approaching GPT-4o's performance. These results highlight Routine's effectiveness in distilling domain-specific tool-usage patterns and enhancing model adaptability to new scenarios. Our experimental results demonstrate that Routine provides a practical and accessible approach to building stable agent workflows, accelerating the deployment and adoption of agent systems in enterprise environments, and advancing the technical vision of AI for Process.

Diffusion Large Language Models (DLLMs) have emerged as a new paradigm of language modeling beyond autoregressive next-token prediction. Thanks to their bidirectional attention mechanism, DLLMs are more capable of capturing the connection of context, and thus show unique advantages in challenges like the famous "reversal curse" or learning under data-constrained scenarios. In addition, taking advantage of their inherent modeling foundations, DLLMs have the great potential of efficient inference with parallel decoding algorithms, which enable multi-token prediction per step. However, the high generation quality often requires the number of decoding steps equal to the sequence length, which performs a one-token-per-step decoding, and existing parallel decoding algorithms, which yield suboptimal decoding paths, bring inference speedup at the cost of non-negligible performance degradation. To overcome this challenge, we introduce Free Draft-and-Verification (FreeDave), a novel fast decoding algorithm tailored for DLLMs that achieves lossless parallel decoding without any model modification or extra modules. Specifically, we propose an algorithm of parallel-decoded candidate generation and verification, which is theoretically guaranteed to use the fewest model forward calls to reproduce the same sequence generated by static decoding when enough computation and memory budget is provided. By extensive evaluations on math reasoning and code generation benchmarks across different DLLMs, FreeDave is proven to boost the inference throughput up to 3.78times without performance degradation.

Fixed-cardinality retrieval injects a constant top-K chunks into the generator regardless of query complexity, causing over-retrieval for narrow queries and under-retrieval for compositional ones. We describe ScoreGate, a lightweight score-space decision mechanism that controls retrieval cardinality at inference time using two scores already produced by the standard pipeline: bi-encoder similarity s_i and cross-encoder reranker score r_i, with no additional model inference calls required. Its core insight is that cross-encoder affirmation can rescue semantically relevant chunks that bi-encoder retrieval ranks poorly due to vocabulary mismatch -- a failure mode unaddressed by fixed-K or single-score thresholding. On MS MARCO (200 dev queries), ScoreGate achieves MRR@10 = 0.401 with 35% fewer retained chunks than Standard Top-K. On an internal benchmark (n=300, Fleiss' kappa=0.87), ScoreGate observed zero false positives (95% CI [96.4%, 100%]) at 97.77-99.34% recall, with 34.8% fewer tokens per query and only 31ms added latency. Results on both MS MARCO and real-world production traffic suggest that adaptive retrieval cardinality can improve retrieval efficiency without degrading retrieval quality.

Agentic systems, AI architectures that autonomously execute multi-step workflows to achieve complex goals, are often built using repeated large language model (LLM) calls for closed-set decision tasks such as routing, shortlisting, gating, and verification. While convenient, this design makes deployments slow and expensive due to cumulative latency and token usage. We propose TabAgent, a framework for replacing generative decision components in closed-set selection tasks with a compact textual-tabular classifier trained on execution traces. TabAgent (i) extracts structured schema, state, and dependency features from trajectories (TabSchema), (ii) augments coverage with schema-aligned synthetic supervision (TabSynth), and (iii) scores candidates with a lightweight classifier (TabHead). On the long-horizon AppWorld benchmark, TabAgent maintains task-level success while eliminating shortlist-time LLM calls, reducing latency by approximately 95% and inference cost by 85-91%. Beyond tool shortlisting, TabAgent generalizes to other agentic decision heads, establishing a paradigm for learned discriminative replacements of generative bottlenecks in production agent architectures.

Systematic reviews and meta-analyses often need numerical data reported only in figures, and extracting them with interactive digitisers usually requires a person to select and calibrate each figure. We present PlotPick, an open-source tool that uses vision-language models (VLMs) to extract tabular data from batches of scientific figures, and we benchmark the kind of model it calls: nine VLMs from four providers on ChartX and six of them on PlotQA, against DePlot, a dedicated chart-to-table model, with every system scored on the same items by numeric F1 (F1 over unlabelled numbers at 5% relative tolerance). On six ChartX chart types (n=300) all nine VLMs outperform DePlot in aggregate, at 79.1-96.0% against 74.3%. The lead comes mainly from box plots, where DePlot scores 24.8% against 64.2-97.3%; pooled over the other five types, seven VLMs keep a lead of 4.7 to 11.6 points and the two weakest do not. On a subset of the PlotQA test split (n=529; 427 horizontal bar charts), scored by a lenient best-series variant of the metric, DePlot reaches 87.0%; the two strongest VLMs are level with it or slightly above it, and four fall 3.8 to 30.3 points below. DePlot was trained on PlotQA's training split. Both benchmarks use synthetic charts, and the metric ignores which series a value belongs to. The application itself was not evaluated: its figure detection, structured output, prompt and default model were not tested, and the one benchmarked model it offers, Claude Haiku 4.5, is one of the four below DePlot on PlotQA. Accuracy on biomedical figures has not been established, and every extracted value must be checked against its source figure. This version corrects version 1, which scored most PlotQA replies against category labels instead of plotted values; its claim that every VLM outperformed DePlot on both benchmarks is withdrawn. PlotPick is available at https://plotpick.streamlit.app/.

Reliable robotic manipulation requires both semantic understanding and precise control in three-dimensional space. However, vision-language-action (VLA) foundations largely inherit visual representations from 2D vision-language pre-training, while spatial supervision is often introduced only during task-specific adaptation. Developing spatial understanding as a transferable foundation-model capability therefore calls for learning from broader robot experience. We present SLA, a Spatial-Language-Action Model that scales implicit spatial alignment to continual pre-training and maintains it during downstream post-training. Initialized from a pretrained pi05 policy, SLA jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. We curate and unify heterogeneous robot datasets to support spatial learning across diverse embodiments, tasks, and scenes. To accommodate this diversity, we introduce feature-dependent gates that adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision. The teacher further exploits partially available depth and camera calibration alongside RGB observations, enabling effective use of heterogeneous geometric annotations. Probabilistic state dropout encourages reliance on visual cues and mitigates overfitting to demonstration trajectories. All auxiliary components are used only during training, allowing direct reuse of pretrained VLA weights and preserving the original inference architecture.

Model routing can reduce the cost of LLM applications by assigning simpler requests to cheaper models without sacrificing the quality of the response. However, most routers treat requests as independent. Stateful agents do not operate under this assumption: a user turn may require multiple model and tool calls, its difficulty depends on the state produced by earlier turns, and changing models can sacrifice reuse of the conversational prompt cache. We study zero-shot, cost-aware routing at the user-turn boundary, systematically varying the routing rule, access to session history, candidate models that differ in size, reasoning effort, or model provider, and prompt-cache behavior. We evaluate on multi-turn versions of SheetCopilot and SpreadsheetBench 2, where persistent workbooks and cross-turn dependencies make isolated routing decisions especially brittle. At a matched offloading rate, giving the router session history improves task pass rate over the same router without history, with gains for every model pair; history-aware routing remains within one point of always using the expensive model. We also found that cost depends on execution dynamics rather than token prices alone: the cheaper model's per-task cost advantage narrows from 6.8x to 3.2x on longer sessions. A closed form cache analysis further shows that downward switches (expensive to cheap) recover their cache cost within two LLM calls, whereas upward switches do not break even under the evaluated conditions. Together, these results show that agent routing is a sequential decision problem: effective policies must account for what happened earlier in the session and for the asymmetric cost of switching models.

Large language models (LLMs) are increasingly deployed as agents in various contexts by providing tools at their disposal. However, LLM agents can exhibit unpredictable behaviors, including taking undesirable and/or unsafe actions. In order to measure the latent propensity of LLM agents for taking illegal actions under an EU legislative context, we introduce EU-Agent-Bench, a verifiable human-curated benchmark that evaluates an agent's alignment with EU legal norms in situations where benign user inputs could lead to unlawful actions. Our benchmark spans scenarios across several categories, including data protection, bias/discrimination, and scientific integrity, with each user request allowing for both compliant and non-compliant execution of the requested actions. Comparing the model's function calls against a rubric exhaustively supported by citations of the relevant legislature, we evaluate the legal compliance of frontier LLMs, and furthermore investigate the compliance effect of providing the relevant legislative excerpts in the agent's system prompt along with explicit instructions to comply. We release a public preview set for the research community, while holding out a private test set to prevent data contamination in evaluating upcoming models. We encourage future work extending agentic safety benchmarks to different legal jurisdictions and to multi-turn and multilingual interactions. We release our code on https://github.com/ilijalichkovski/eu-agent-bench{this URL}.

Agentic systems enable the intelligent use of research tooling, augmenting a researcher's ability to investigate and propose novel solutions to existing problems. Within Additive Manufacturing (AM), alloy discovery remains a complex challenge, often requiring expertise in the various domains of materials science, thermodynamic simulations, and experimental analysis. Large Language Model (LLM) enabled agents can facilitate this endeavor by utilizing their extensive knowledge base to dispatch tool calls via Model Context Protocol (MCP) to perform actions such as Thermo-Calc property diagram calculations and lack of fusion process map generation. In addition, the multi-agent system developed in this work is able to effectively reason through complex user prompts and provide analysis on the printability of proposed alloys. These agents can dynamically adjust their task trajectory to the outcomes of tool call results, effectively enabling autonomous decision-making in practical environments. This work aims to utilize LLM enabled agents to automate and accelerate the task of alloy discovery within the field of additive manufacturing and showcase the benefits of adopting this multi-agent system.

LLM-agent services repeatedly execute small deterministic transitions between model and tool calls: route an outcome, update state, and emit the next effect. We ask when this control path exposes enough concurrent work for GPU execution, and what changes when a GPU-computed route decision remains on device. We formalize the ready-cohort boundary using fixed-partition share F, exact offline share P*, local upper bound U, and online achieved share A. Under zero service time, unlimited capacity, and equal relative launch deadlines, a specialized dynamic program computes P* exactly. In a stationary Poisson replay of one pinned 851-session public trace panel, the primary condition at 100,000 target active sessions, K=256, and a 50 ms launch deadline gives F=30.19%, P*=43.00%, and U=45.85%. Exact packing recovers 81.83% of the opportunity lost at fixed window boundaries. The outcome-derived route key is a conditioning proxy, not proof of executable identity. A separate mechanism study keeps a GPU-computed binary decision on device instead of returning four bytes to the host and redispatching. Across four named GPU placements, the device-resident path is faster in all 36 configurations; within-placement row-median ratios range from 1.19x to 2.39x. Across both admissible mechanisms, all 14,557,440 tested batched invocations match a separately implemented host oracle. A fixed nested device graph that removes no host decision is slower in all 60 configurations across five placements. Together, the studies establish two measurable gates for GPU agent control: deadline-feasible cohort supply and observation placement. A joined finite online runtime is required to measure A, CPU displacement, and service-level benefit.

We present the first systematic approach to static and dynamic taint analysis for Graph APIs focusing on broken access control. The approach comprises the following. We taint nodes in the Graph API if they represent data requiring specific privileges in order to be retrieved or manipulated, and identify API calls which are related to sources and sinks. Then, we statically analyze whether tainted information flow between API source and sink calls occurs. To this end, we model the API calls using graph transformation rules. We subsequently use critical pair analysis to automatically analyze potential dependencies between rules representing source calls and rules representing sink calls. We distinguish direct from indirect tainted information flow and argue under which conditions the CPA is able to detect not only direct, but also indirect tainted flow. The static taint analysis (i) identifies flows that need to be further reviewed, since tainted nodes may be created by an API call and used or manipulated by another API call later without having the necessary privileges, and (ii) can be used to systematically design dynamic security tests for broken access control. The dynamic taint analysis checks if potential broken access control risks detected during the static taint analysis really occur. We apply the approach to a part of the GitHub GraphQL API. The application illustrates that our analysis supports the detection of two types of broken access control systematically: the case where users of the API may not be able to access or manipulate information, although they should be able to do so; and the case where users (or attackers) of the API may be able to access/manipulate information that they should not.

Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its comprehensive architecture by analyzing the publicly available TypeScript source code and further comparing it with OpenClaw, an independent open-source AI agent system that answers many of the same design questions from a different deployment context. Our analysis identifies five human values, philosophies, and needs that motivate the architecture (human decision authority, safety and security, reliable execution, capability amplification, and contextual adaptability) and traces them through thirteen design principles to specific implementation choices. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. Most of the code, however, lives in the systems around this loop: a permission system with seven modes and an ML-based classifier, a five-layer compaction pipeline for context management, four extensibility mechanisms (MCP, plugins, skills, and hooks), a subagent delegation mechanism with worktree isolation, and append-oriented session storage. A comparison with OpenClaw, a multi-channel personal assistant gateway, shows that the same recurring design questions produce different architectural answers when the deployment context changes: from per-action safety classification to perimeter-level access control, from a single CLI loop to an embedded runtime within a gateway control plane, and from context-window extensions to gateway-wide capability registration. We finally identify six open design directions for future agent systems, grounded in recent empirical, architectural, and policy literature.

LLM agents increasingly rely on community-contributed skills that expand an agent's operational capability set. We study a core safety problem in agentic AI systems: whether individually safe skills can compose into unsafe installed skill sets. We present SkillReact, a compositional security measurement framework with three components: a deterministic static-composition benchmark, a two-rater LLM-assisted human-adjudication pipeline, and an action-based exploitability harness. On 1,520 ClawHub skills, 651 pass individual inspection and form 211,575 pairs; the benchmark flags 22.25% of these as structural candidates. We treat this raw rate as a recall-oriented scanner ceiling and calibrate it against human judgment: in a pattern-stratified audit, roughly one in five flagged pair-pattern hits survives as a real compositional risk (population-weighted validity 18.2%, our headline result), implying about 14K genuine risk memberships in a single registry that per-skill scanning misses by construction, since every pair is individually safe. An action-based harness then probes when these candidates become model-issued tool calls, and finds realization gated by host-model disposition: on an anchor-conditioned dropper subset, Haiku-4-5 issues the dropper-stage tool call on all 39 direct-prompt trials (36 of them the full download-then-execute chain, 3 download-only), Opus-4-7 stops at the download, and Sonnet-4-6 refuses outright. A control that holds the request fixed and varies only the installed skills finds compliance highest with no skills installed: a composition fixes which capabilities are reachable, while the host model decides whether to use them. Together these motivate install-time compositional checks and capability isolation as complements to per-skill scanning.

AI agents increasingly act through external tools: they query databases, execute shell commands, read and write files, and send network requests. Yet in most current agent stacks, model-generated tool calls are handed to the execution layer with no framework-agnostic control point in between. Post-execution observability can record these actions, but it cannot stop them before side effects occur. We present AEGIS, a pre-execution firewall and audit layer for AI agents. AEGIS interposes on the tool-execution path and applies a three-stage pipeline: (i) deep string extraction from tool arguments, (ii) content-first risk scanning, and (iii) composable policy validation. High-risk calls can be held for human approval, and all decisions are recorded in a tamper-evident audit trail based on Ed25519 signatures and SHA-256 hash chaining. In the current implementation, AEGIS supports 14 agent frameworks across Python, JavaScript, and Go with lightweight integration. On a curated suite of 48 attackinstances, AEGIS blocks all attacks in the suite before execution; on 500 benign tool calls, it yields a 1.2% false positive rate; and across 1,000 consecutive interceptions, it adds 8.3 ms median latency. The live demo will show end-to-end interception of benign, malicious, and human-escalated tool calls, allowing attendees to observe real-time blocking, approval workflows, and audit-trail generation. These results suggest that pre-execution mediation for AI agents can be practical, low-overhead, and directly deployable.

Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to improve their internal reasoning ability by integrating external tools. However, models employing TIR often display suboptimal behaviors, such as insufficient or excessive tool usage and overthinking after tool calls. The challenge of incentivizing LLMs to perform TIR efficiently and accurately, while stabilizing the reasoning process, remains an open question. In this paper, we start by exploring the impact of tool calls on model reasoning from the perspective of information entropy. Our findings indicate that tool call results lead to a distinct change in the information entropy of subsequent reasoning, with the overall entropy of the reasoning chain varying based on the number of tool calls. Building on these insights, we propose Tool-Light, a framework designed to encourage LLMs to perform TIR efficiently and accurately. Our framework includes dataset construction and multi-stage fine-tuning. For dataset construction, we employ continuous self-evolved sampling using the fine-tuned model, integrating both vanilla sampling and entropy-guided sampling. Besides, we establish strict criteria for selecting positive-negative pairs during sampling. The training process involves a two-stage approach, comprising Supervised Fine-Tuning (SFT) and Self-Evolved Direct Preference Optimization (DPO). Experimental results on 10 datasets demonstrate the effectiveness of Tool-Light, significantly improving the model's efficiency in executing TIR tasks.

Artificial Intelligence has become a double edged sword in modern society being both a boon and a bane. While it empowers individuals it also enables malicious actors to perpetrate scams such as fraudulent phone calls and user impersonations. This growing threat necessitates a robust system to protect individuals In this paper we introduce a novel real time fraud detection mechanism using Retrieval Augmented Generation technology to address this challenge on two fronts. First our system incorporates a continuously updating policy checking feature that transcribes phone calls in real time and uses RAG based models to verify that the caller is not soliciting private information thus ensuring transparency and the authenticity of the conversation. Second we implement a real time user impersonation check with a two step verification process to confirm the callers identity ensuring accountability. A key innovation of our system is the ability to update policies without retraining the entire model enhancing its adaptability. We validated our RAG based approach using synthetic call recordings achieving an accuracy of 97.98 percent and an F1score of 97.44 percent with 100 calls outperforming state of the art methods. This robust and flexible fraud detection system is well suited for real world deployment.

Multi-agent systems (MAS) built from large language models coordinate specialized agents to tackle complex tasks, but effective workflows are difficult to design in advance. Test-time evolution refines workflows using execution feedback, yet broad revisions can disturb useful components, while re-executing unchanged requests can incur redundant computation. Inspired by the interplay of inheritance and selection in biological evolution, we introduce Inherit-MAS, which makes inheritance explicit at the workflow and execution levels. A meta-model first synthesizes a workflow of worker agents with declared roles, communication inputs, and tool permissions, and a separately prompted judge scores each executed candidate and diagnoses its deficiencies. In ordinary refinement rounds, workflow inheritance starts from the latest completed candidate, may discard removable nodes judged unhelpful, and applies a validated edit to address the diagnosed deficiency. When the new candidate executes, execution inheritance inherits eligible stored results only if the complete resolved request and execution context match, avoiding redundant model and tool calls. With GPT-4o-mini workers, Inherit-MAS achieves 55.4\% completion on WorkBench and 49.7\% joint F1 on HotpotQA FullWiki, outperforming EvoAgent, EvoMAS, and TacoMAS. With Qwen3-32B workers, it also exceeds these evolving-MAS baselines on both benchmarks. Compared with rerunning the same controller with execution inheritance disabled, execution inheritance reduces worker-token usage by 29.1\% on WorkBench and 34.6\% on HotpotQA, and total token usage by 5.3\% and 18.1\%.

Large language models (LLMs) have recently shown tremendous promise in serving as the backbone to agentic systems, as demonstrated by their performance in multi-faceted, challenging benchmarks like SWE-Bench and Agent-Bench. However, to realize the true potential of LLMs as autonomous agents, they must learn to identify, call, and interact with external tools and application program interfaces (APIs) to complete complex tasks. These tasks together are termed function calling. Endowing LLMs with function calling abilities leads to a myriad of advantages, such as access to current and domain-specific information in databases and knowledge sources, and the ability to outsource tasks that can be reliably performed by tools, e.g., a Python interpreter or calculator. While there has been significant progress in function calling with LLMs, there is still a dearth of open models that perform on par with proprietary LLMs like GPT, Claude, and Gemini. Therefore, in this work, we introduce the GRANITE-20B-FUNCTIONCALLING model under an Apache 2.0 license. The model is trained using a multi-task training approach on seven fundamental tasks encompassed in function calling, those being Nested Function Calling, Function Chaining, Parallel Functions, Function Name Detection, Parameter-Value Pair Detection, Next-Best Function, and Response Generation. We present a comprehensive evaluation on multiple out-of-domain datasets comparing GRANITE-20B-FUNCTIONCALLING to more than 15 other best proprietary and open models. GRANITE-20B-FUNCTIONCALLING provides the best performance among all open models on the Berkeley Function Calling Leaderboard and fourth overall. As a result of the diverse tasks and datasets used for training our model, we show that GRANITE-20B-FUNCTIONCALLING has better generalizability on multiple tasks in seven different evaluation datasets.

Recent studies suggest that with sufficiently wide models, most SGD solutions can, up to permutation, converge into the same basin. This phenomenon, known as the model re-basin regime, has significant implications for model averaging by ensuring the linear mode connectivity. However, current re-basin strategies are ineffective in many scenarios due to a lack of comprehensive understanding of underlying mechanisms. Addressing this gap, this paper provides novel insights into understanding and improving the standard practice. Firstly, we decompose re-normalization into rescaling and reshift, uncovering that rescaling plays a crucial role in re-normalization while re-basin performance is sensitive to shifts in model activation. The finding calls for a more nuanced handling of the activation shift. Secondly, we identify that the merged model suffers from the issue of activation collapse and magnitude collapse. Varying the learning rate, weight decay, and initialization method can mitigate the issues and improve model performance. Lastly, we propose a new perspective to unify the re-basin and pruning, under which a lightweight yet effective post-pruning technique is derived, which can significantly improve the model performance after pruning. Our implementation is available at https://github.com/XingyuQu/rethink-re-basin.

This study investigates the automated classification of Calls to Action (CTAs) within the 2021 German Instagram election campaign to advance the understanding of mobilization in social media contexts. We analyzed over 2,208 Instagram stories and 712 posts using fine-tuned BERT models and OpenAI's GPT-4 models. The fine-tuned BERT model incorporating synthetic training data achieved a macro F1 score of 0.93, demonstrating a robust classification performance. Our analysis revealed that 49.58% of Instagram posts and 10.64% of stories contained CTAs, highlighting significant differences in mobilization strategies between these content types. Additionally, we found that FDP and the Greens had the highest prevalence of CTAs in posts, whereas CDU and CSU led in story CTAs.

Language understanding in the brain is context-dependent, varying across experimental stimuli and individuals, which makes it difficult to build computational models that generalize across both. This calls for a foundation model of language-evoked brain activity that can capture shared structure while adapting efficiently to new participants and inputs. We introduce RABBiT (Rapidly Adaptive BOLD foundation model via BraIn-Tuning), a compact audio-to-fMRI encoder designed for accurate zero- and few-shot prediction. A comprehensive evaluation on 324 participants across multiple unseen fMRI datasets shows that RABBiT enables accurate zero-shot prediction of fMRI responses to natural speech across auditory and language-selective regions, surpassing the SOTA foundation model for fMRI and predictions based on group averages. With as little as 10 minutes of participant-specific data, RABBiT further improves performance via parameter-efficient tuning, substantially outperforming per-participant linear models. RABBiT's performance is driven by two key innovations: (1) learned region-specific attention, and (2) a decomposition of brain responses into shared and subject-specific components, combined with a brain-tuned speech backbone. In addition to supporting strong predictive accuracy, the structured, region-specific representations that RABBiT learns enable interpretability. By eliminating the need for extensive per-participant data and model fitting, RABBiT enables scalable population-level analyses of language in the human brain. We make the code available at https://github.com/bridge-ai-neuro/rabbit.

Tool-augmented LLM agents commonly rely on step-wise atomic tool calls, where each invocation, observation, and value transfer is exposed in the main reasoning trace. This creates an execution-granularity mismatch: locally deterministic tool workflows are unfolded into repeated model-visible decisions, consuming context and forcing the model to manage low-level dataflow in the trace. We introduce HyperTool, a unified executable MCP-style tool interface that changes the model-visible unit of tool execution. A model invokes HyperTool with a code block that can call existing tools through their original schemas, manipulate returned values, and pass intermediate results locally, folding deterministic tool subroutines into a single outer call. To train models to use this interface, we synthesize HyperTool-format trajectories from cross-tool compositional tasks and verify them in real MCP environments. On MCP-Universe, HyperTool improves average accuracy from 15.69\% to 35.29\% on Qwen3-32B and from 9.93\% to 33.33\% on Qwen3-8B, and surpass GPT-OSS and Kimi-k2.5 on average accuracy, showing that our HyperTool can substantially improve multi-step tool use.

Large Language Model (LLM) agents rely on external tools to perform multistage tasks. Existing agent frameworks typically assume that tool calls are atomic and return binary success or failure signals. However, real-world systems exhibit non-atomic behaviors such as timeouts after dispatch, delayed visibility, and partial state updates. These mismatches lead to reliability issues including duplicate actions, task success, and unnecessary tool executions. A lightweight, verification-aware tool wrapper is introduced that augments tool calls with postcondition verification, verify-before-retry logic, and idempotency keys. The approach is evaluated in a controlled simulated environment with injected non-atomic failures across multiple task templates. The results demonstrate that the proposed method significantly reduces duplicate actions, while maintaining comparable task success rates. Overall, the findings suggest that strengthening tool interaction semantics is a promising direction for improving LLM agent reliability without requiring modifications to the underlying language model.

Predicting stock market prices following corporate earnings calls remains a significant challenge for investors and researchers alike, requiring innovative approaches that can process diverse information sources. This study investigates the impact of corporate earnings calls on stock prices by introducing a multi-modal predictive model. We leverage textual data from earnings call transcripts, along with images and tables from accompanying presentations, to forecast stock price movements on the trading day immediately following these calls. To facilitate this research, we developed the MiMIC (Multi-Modal Indian Earnings Calls) dataset, encompassing companies representing the Nifty 50, Nifty MidCap 50, and Nifty Small 50 indices. The dataset includes earnings call transcripts, presentations, fundamentals, technical indicators, and subsequent stock prices. We present a multimodal analytical framework that integrates quantitative variables with predictive signals derived from textual and visual modalities, thereby enabling a holistic approach to feature representation and analysis. This multi-modal approach demonstrates the potential for integrating diverse information sources to enhance financial forecasting accuracy. To promote further research in computational economics, we have made the MiMIC dataset publicly available under the CC-NC-SA-4.0 licence. Our work contributes to the growing body of literature on market reactions to corporate communications and highlights the efficacy of multi-modal machine learning techniques in financial analysis.

Compound AI systems that combine multiple LLM calls, such as self-refine and multi-agent-debate, achieve strong performance on many AI tasks. We address a core question in optimizing compound systems: for each LLM call or module in the system, how should one decide which LLM to use? We show that these LLM choices have a large effect on quality, but the search space is exponential. We propose LLMSelector, an efficient framework for model selection in compound systems, which leverages two key empirical insights: (i) end-to-end performance is often monotonic in how well each module performs, with all other modules held fixed, and (ii) per-module performance can be estimated accurately by an LLM. Building upon these insights, LLMSelector iteratively selects one module and allocates to it the model with the highest module-wise performance, as estimated by an LLM, until no further gain is possible. LLMSelector is applicable to any compound system with a bounded number of modules, and its number of API calls scales linearly with the number of modules, achieving high-quality model allocation both empirically and theoretically. Experiments with popular compound systems such as multi-agent debate and self-refine using LLMs such as GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 show that LLMSelector confers 5%-70% accuracy gains compared to using the same LLM for all modules.

Modern TTS systems increasingly generate synthetic speech at scale for diverse users. This setting calls for content-level provenance that can verify the origin of released speech and attribute it to the requesting user, which generative watermarking can support by embedding multi-bit identifiers directly into synthesized speech. Once released, however, speech may undergo heterogeneous learned transformations during distribution and editing, with reconstruction objectives that can preserve speech utility while affecting watermark recoverability differently. We find that no single watermark carrier remains consistently reliable across reconstruction models, as its survival depends jointly on the embedded structure, reconstruction mechanism, and observation representation. To this end, we propose Thrive, a multi-bit generative speech watermarking framework for modern autoregressive TTS, covering both discrete-token and continuous-representation generation under reconstruction attacks. Specifically, Rise synchronizes watermark injection into intermediate representations with its continued integration into subsequent generation, while Care combines waveform and spectral experts using bit-wise reliability selection. Experiments on both autoregressive paradigms show that Thrive preserves synthesis fidelity, achieves 87.6% average recovery accuracy under reconstruction attacks, and supports source attribution over candidate sets of up to 10,000 identities.

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

Large Language Models (LLMs) excel at generating fluent text but struggle to enforce external constraints because they generate tokens sequentially without explicit control mechanisms. GenCP addresses this limitation by combining LLM predictions with Constraint Programming (CP) reasoning, formulating text generation as a Constraint Satisfaction Problem (CSP). In this paper, we improve GenCP by integrating Masked Language Models (MLMs) for domain generation, which allows bidirectional constraint propagation that leverages both past and future tokens. This integration bridges the gap between token-level prediction and structured constraint enforcement, leading to more reliable and constraint-aware text generation. Our evaluation on COLLIE benchmarks demonstrates that incorporating domain preview via MLM calls significantly improves GenCP's performance. Although this approach incurs additional MLM calls and, in some cases, increased backtracking, the overall effect is a more efficient use of LLM inferences and an enhanced ability to generate feasible and meaningful solutions, particularly in tasks with strict content constraints.

Large language models (LLMs) have become increasingly capable of following instructions and complex reasoning, making prompting a flexible interface for adapting models without parameter updates. Yet prompt design remains labor-intensive and highly sensitive to formatting, phrasing, and instruction order, motivating automated prompt optimization methods that reduce manual effort while preserving inference-time flexibility. However, existing methods often search over prompt candidates or use fixed critique-refine pipelines driven by individual examples or small batches, limiting their ability to capture systematic error patterns and make targeted edits grounded in failure history. We propose Reflective Prompt Tuning (RPT), a framework that uses LLM function calling to simulate the iterative workflow of human prompt engineers. An LLM optimizer calls a diagnostic function that evaluates the target model over an entire optimization set, summarizes recurring failure modes, and returns a structured diagnostic report. The optimizer uses this report, together with an accumulated memory of prior reports, to revise the prompt for the next iteration. RPT further supports confidence-aware optimization by using calibration signals in diagnostic feedback and final prompt selection. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points, remains competitive with state of the art, and improves confidence calibration. Our analyses show that RPT is especially effective on multi-hop and mathematical reasoning, producing targeted prompt revisions that align with diagnosed failure patterns and lead to gains in task performance and calibration.

Given large-scale source code datasets available in open-source projects and advanced large language models, recent code models have been proposed to address a series of critical software engineering tasks, such as program repair and code completion. The training data of the code models come from various sources, not only the publicly available source code, e.g., open-source projects on GitHub but also the private data such as the confidential source code from companies, which may contain sensitive information (for example, SSH keys and personal information). As a result, the use of these code models may raise new privacy concerns. In this paper, we focus on a critical yet not well-explored question on using code models: what is the risk of membership information leakage in code models? Membership information leakage refers to the risk that an attacker can infer whether a given data point is included in (i.e., a member of) the training data. To answer this question, we propose Gotcha, a novel membership inference attack method specifically for code models. We investigate the membership leakage risk of code models. Our results reveal a worrying fact that the risk of membership leakage is high: although the previous attack methods are close to random guessing, Gotcha can predict the data membership with a high true positive rate of 0.95 and a low false positive rate of 0.10. We also show that the attacker's knowledge of the victim model (e.g., the model architecture and the pre-training data) impacts the success rate of attacks. Further analysis demonstrates that changing the decoding strategy can mitigate the risk of membership leakage. This study calls for more attention to understanding the privacy of code models and developing more effective countermeasures against such attacks.

Weakly supervised object localization (WSOL) remains challenging when learning object localization models from image category labels. Conventional methods that discriminatively train activation models ignore representative yet less discriminative object parts. In this study, we propose a generative prompt model (GenPromp), defining the first generative pipeline to localize less discriminative object parts by formulating WSOL as a conditional image denoising procedure. During training, GenPromp converts image category labels to learnable prompt embeddings which are fed to a generative model to conditionally recover the input image with noise and learn representative embeddings. During inference, enPromp combines the representative embeddings with discriminative embeddings (queried from an off-the-shelf vision-language model) for both representative and discriminative capacity. The combined embeddings are finally used to generate multi-scale high-quality attention maps, which facilitate localizing full object extent. Experiments on CUB-200-2011 and ILSVRC show that GenPromp respectively outperforms the best discriminative models by 5.2% and 5.6% (Top-1 Loc), setting a solid baseline for WSOL with the generative model. Code is available at https://github.com/callsys/GenPromp.

Every fiscal quarter, companies hold earnings calls in which company executives respond to questions from analysts. After these calls, analysts often change their price target recommendations, which are used in equity research reports to help investors make decisions. In this paper, we examine analysts' decision making behavior as it pertains to the language content of earnings calls. We identify a set of 20 pragmatic features of analysts' questions which we correlate with analysts' pre-call investor recommendations. We also analyze the degree to which semantic and pragmatic features from an earnings call complement market data in predicting analysts' post-call changes in price targets. Our results show that earnings calls are moderately predictive of analysts' decisions even though these decisions are influenced by a number of other factors including private communication with company executives and market conditions. A breakdown of model errors indicates disparate performance on calls from different market sectors.

LLMs' capabilities are enhanced by using function calls to integrate various data sources or API results into the context window. Typical tools include search, web crawlers, maps, financial data, file systems, and browser usage, etc. Integrating these data sources or functions requires a standardized method. The Model Context Protocol (MCP) provides a standardized way to supply context to LLMs. However, the evaluation of LLMs and AI Agents' MCP tool use abilities suffer from several issues. First, there's a lack of comprehensive datasets or benchmarks to evaluate various MCP tools. Second, the diverse formats of response from MCP tool call execution further increase the difficulty of evaluation. Additionally, unlike existing tool-use benchmarks with high success rates in functions like programming and math functions, the success rate of real-world MCP tool is not guaranteed and varies across different MCP servers. Furthermore, the LLMs' context window also limits the number of available tools that can be called in a single run, because the textual descriptions of tool and the parameters have long token length for an LLM to process all at once. To help address the challenges of evaluating LLMs' performance on calling MCP tools, we propose MCPToolBench++, a large-scale, multi-domain AI Agent tool use benchmark. As of July 2025, this benchmark is build upon marketplace of over 4k MCP servers from more than 40 categories, collected from the MCP marketplaces and GitHub communities. The datasets consist of both single-step and multi-step tool calls across different categories. We evaluated SOTA LLMs with agentic abilities on this benchmark and reported the results.

Detecting evasive answers in earnings calls is critical for financial transparency, yet progress is hindered by the lack of large-scale benchmarks. We introduce EvasionBench, comprising 30,000 training samples and 1,000 human-annotated test samples (Cohen's Kappa 0.835) across three evasion levels. Our key contribution is a multi-model annotation framework leveraging a core insight: disagreement between frontier LLMs signals hard examples most valuable for training. We mine boundary cases where two strong annotators conflict, using a judge to resolve labels. This approach outperforms single-model distillation by 2.4 percent, with judge-resolved samples improving generalization despite higher training loss (0.421 vs 0.393) - evidence that disagreement mining acts as implicit regularization. Our trained model Eva-4B (4B parameters) achieves 81.3 percent accuracy, outperforming its base by 25 percentage points and approaching frontier LLM performance at a fraction of inference cost.

Long agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it with fixed-interval compaction triggered at a token threshold. Such triggers pay no heed to trajectory structure, risking discard of partial results mid-derivation or mid-search. We propose SelfCompact, a scaffold that allows the model itself to decide when and how to compact. Specifically, it pairs two inference-time elements: (i) a compaction tool the model invokes to summarize the accumulated context, and (ii) a lightweight rubric specifying when to fire (a sub-task has resolved, or the trajectory is converging) and when to suppress (mid-derivation, or when stuck). Both are needed. The tool alone is unevenly used across open-weight models, often invoked at unhelpful moments or not at all; the rubric alone cannot act. Together, they elicit effective adaptive compaction without any fine-tuning or external supervision. We present empirical results on six benchmarks (competitive math and agentic search) and seven models. Our results show that SelfCompact matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 18.1 points on math and 5-9 points on agentic search at 30-70% lower per-question cost. Our results expose a meta-cognitive gap: although unprompted models cannot reliably tell when their own context is rotting, a lightweight rubric closes this gap, reframing when to compact as a capability that scaffolds can supply without training.

A voice agent has to decide, at every pause, whether the user has finished speaking. Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait. Open semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language. We release TamilEOT: 18,485 labelled turn boundaries cut from 116 real Tamil telephone conversations, and two audio-only detectors fine-tuned from Smart Turn v3. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both models run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Rule-derived labels, checked against a blind human listening pass, were right 95.9% of the time on the positive class and 44.4% on the negative class, which is below chance, because the rule answered a different question than the model is asked. Replacing them with an audio-LLM labeller measured at 97.5% human agreement cost US$5.69. Of every training lever we measured, only encoder capacity moved the result; three runs at identical config and seed span 0.87 accuracy points, which is the floor below which none of our other deltas mean anything. Replaying the same labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries are never surfaced to the model at all. Data, weights, code and every negative result are public.

Transcriptions of phone calls are of significant value across diverse fields, such as sales, customer service, healthcare, and law enforcement. Nevertheless, the analysis of these recorded conversations can be an arduous and time-intensive process, especially when dealing with extended or multifaceted dialogues. In this work, we propose a novel method, GPT-distilled Calls Segmentation and Tagging (GPT-Calls), for efficient and accurate call segmentation and topic extraction. GPT-Calls is composed of offline and online phases. The offline phase is applied once to a given list of topics and involves generating a distribution of synthetic sentences for each topic using a GPT model and extracting anchor vectors. The online phase is applied to every call separately and scores the similarity between the transcripted conversation and the topic anchors found in the offline phase. Then, time domain analysis is applied to the similarity scores to group utterances into segments and tag them with topics. The proposed paradigm provides an accurate and efficient method for call segmentation and topic extraction that does not require labeled data, thus making it a versatile approach applicable to various domains. Our algorithm operates in production under Dynamics 365 Sales Conversation Intelligence, and our research is based on real sales conversations gathered from various Dynamics 365 Sales tenants.

Large Language Models (LLMs) have been gaining increasing attention and demonstrated promising performance across a variety of Software Engineering (SE) tasks, such as Automated Program Repair (APR), code summarization, and code completion. For example, ChatGPT, the latest black-box LLM, has been investigated by numerous recent research studies and has shown impressive performance in various tasks. However, there exists a potential risk of data leakage since these LLMs are usually close-sourced with unknown specific training details, e.g., pre-training datasets. In this paper, we seek to review the bug-fixing capabilities of ChatGPT on a clean APR benchmark with different research objectives. We first introduce {\benchmark}, a new benchmark with buggy and the corresponding fixed programs from competitive programming problems starting from 2023, after the training cutoff point of ChatGPT. The results on {\benchmark} show that ChatGPT is able to fix 109 out of 151 buggy programs using the basic prompt within 35 independent rounds, outperforming state-of-the-art LLMs CodeT5 and PLBART by 27.5\% and 62.4\% prediction accuracy. We also investigate the impact of three types of prompts, i.e., problem description, error feedback, and bug localization, leading to additional 34 fixed bugs. Besides, we provide additional discussion from the interactive nature of ChatGPT to illustrate the capacity of a dialog-based repair workflow with 9 additional fixed bugs. Inspired by the findings, we further pinpoint various challenges and opportunities for advanced SE study equipped with such LLMs (e.g.,~ChatGPT) in the near future. More importantly, our work calls for more research on the reevaluation of the achievements obtained by existing black-box LLMs across various SE tasks, not limited to ChatGPT on APR.

To reduce development overhead and enable seamless integration between potential components comprising any given generative AI application, the Model Context Protocol (MCP) (Anthropic, 2024) has recently been released and subsequently widely adopted. The MCP is an open protocol that standardizes API calls to large language models (LLMs), data sources, and agentic tools. By connecting multiple MCP servers, each defined with a set of tools, resources, and prompts, users are able to define automated workflows fully driven by LLMs. However, we show that the current MCP design carries a wide range of security risks for end users. In particular, we demonstrate that industry-leading LLMs may be coerced into using MCP tools to compromise an AI developer's system through various attacks, such as malicious code execution, remote access control, and credential theft. To proactively mitigate these and related attacks, we introduce a safety auditing tool, MCPSafetyScanner, the first agentic tool to assess the security of an arbitrary MCP server. MCPScanner uses several agents to (a) automatically determine adversarial samples given an MCP server's tools and resources; (b) search for related vulnerabilities and remediations based on those samples; and (c) generate a security report detailing all findings. Our work highlights serious security issues with general-purpose agentic workflows while also providing a proactive tool to audit MCP server safety and address detected vulnerabilities before deployment. The described MCP server auditing tool, MCPSafetyScanner, is freely available at: https://github.com/johnhalloran321/mcpSafetyScanner

Augmenting Large Language Models (LLMs) with external tools enables them to execute complex, multi-step tasks. However, tool learning is hampered by the static synthetic data pipelines where data generation and model training are executed as two separate, non-interactive processes. This approach fails to adaptively focus on a model's specific weaknesses and allows noisy labels to persist, degrading training efficiency. We introduce LoopTool, a fully automated, model-aware data evolution framework that closes this loop by tightly integrating data synthesis and model training. LoopTool iteratively refines both the data and the model through three synergistic modules: (1) Greedy Capability Probing (GCP) diagnoses the model's mastered and failed capabilities; (2) Judgement-Guided Label Verification (JGLV) uses an open-source judge model to find and correct annotation errors, progressively purifying the dataset; and (3) Error-Driven Data Expansion (EDDE) generates new, challenging samples based on identified failures. This closed-loop process operates within a cost-effective, open-source ecosystem, eliminating dependence on expensive closed-source APIs. Experiments show that our 8B model trained with LoopTool significantly surpasses its 32B data generator and achieves new state-of-the-art results on the BFCL-v3 and ACEBench benchmarks for its scale. Our work demonstrates that closed-loop, self-refining data pipelines can dramatically enhance the tool-use capabilities of LLMs.

This technical white paper introduces the Interactive Agents Call Tree (IACT), a computational model designed to address the limitations of static, hard-coded agent workflows. Unlike traditional systems that require pre-defined graphs or specialized programming, IACT operates as a general-purpose autonomous system driven purely by user dialogue. Given a high-level objective, the system autonomously grows a dynamic, recursive agent topology incrementally tailored to the problem's structure. This allows it to scale its organizational complexity to match open-ended tasks. To mitigate the error propagation inherent in unidirectional function calls, IACT introduces interactional redundancy by replacing rigid invocations with bidirectional, stateful dialogues. This mechanism enables runtime error correction and ambiguity resolution. We describe the architecture, design principles, and practical lessons behind the production deployment of this model in the kragent.ai system, presenting qualitative evidence from real-world workflows rather than exhaustive benchmark results.

The Model Context Protocol (MCP) (MCP Community, 2025) has emerged as a widely used framework for enabling LLM-based agents to communicate with external tools and services. The original MCP implementation (Anthropic, 2024) relies on a Large Language Model (LLM) to decompose tasks and issue instructions to servers. In particular, the agents, models, and servers are stateless and do not have access to a global context. However, in tasks involving LLM-driven coordination, it is natural that a Shared Context Store (SCS) could improve the efficiency and coherence of multi-agent workflows by reducing redundancy and enabling knowledge transfer between servers. Thus, in this work, we design and assess the performance of a Context-Aware MCP (CA-MCP) that offloads execution logic to specialized MCP servers that read from and write to a shared context memory, allowing them to coordinate more autonomously in real time. In this design, context management serves as the central mechanism that maintains continuity across task executions by tracking intermediate states and shared variables, thereby enabling persistent collaboration among agents without repeated prompting. We present experiments showing that the CA-MCP can outperform the traditional MCP by reducing the number of LLM calls required for complex tasks and decreasing the frequency of response failures when task conditions are not satisfied. In particular, we conducted experiments on the TravelPlanner (Yang et al., 2024) and REALM-Bench (Geng & Chang, 2025) benchmark datasets and observed statistically significant results indicating the potential advantages of incorporating a shared context store via CA-MCP in LLM-driven multi-agent systems.

Speech and audio LLMs are evaluated by comparing waveform predictions with predictions from an automatic speech recognition (ASR) transcript. For fixed closed-set tasks, this conflates acoustic evidence with the need to invoke a generative audio model. We estimate incremental call value with matched selectors sharing pre-call evidence. Each policy may retain the transcript label, use a local encoder, or invoke a generative model; matched control removes generative actions but preserves pre-call evidence and development selection. On VocalSound, transcript-only accuracy is 0.296, while supervised CLAP and WavLM controls reach 0.850 and 0.854 without calls. Full selector reaches 0.925 at 12.5% calls versus 0.921 for matched No-call selector (difference 0.004; 95% CI [-0.025, 0.033]). Thus, results do not show a call gain after transcript and encoder evidence are available. Relevant quantity is incremental accuracy from allowing calls, not the waveform-transcript gap.

In this paper, we present an update to the NISQA speech quality prediction model that is focused on distortions that occur in communication networks. In contrast to the previous version, the model is trained end-to-end and the time-dependency modelling and time-pooling is achieved through a Self-Attention mechanism. Besides overall speech quality, the model also predicts the four speech quality dimensions Noisiness, Coloration, Discontinuity, and Loudness, and in this way gives more insight into the cause of a quality degradation. Furthermore, new datasets with over 13,000 speech files were created for training and validation of the model. The model was finally tested on a new, live-talking test dataset that contains recordings of real telephone calls. Overall, NISQA was trained and evaluated on 81 datasets from different sources and showed to provide reliable predictions also for unknown speech samples. The code, model weights, and datasets are open-sourced.

An agentic system issues several structurally different kinds of LLM calls. It routes intent, classifies actions, grounds language in a device registry, plans multi-agent pipelines and writes the Python code those pipelines run. The difficulty of these call sites varies by an order of magnitude, yet in practice a single model, chosen for the hardest site, serves all of them. In this work, we evaluate 9 models from 0.8B to a frontier hosted model across the five call sites of a deployed open-source home-automation framework (Wactorz), using its unmodified production prompts and two real Home Assistant installations (280 cases, 2520 scored calls). We find that capability is not ordered the same way at every site, and that larger models are not uniformly better: one 4B model is worse than its 2B sibling at grounded actuation. Paired testing shows the best local model to be statistically indistinguishable from both hosted models at four of five sites. Only code generation separates them, against a small hosted model (p = 0.039) as well as a frontier one (p = 0.002). Aggregate accuracy also hides a safety failure specific to actuation, where small models resolve the accuracy/refusal trade-off in degenerate ways: one model (Gemma4 E2B) actuates on 87.2% of requests for devices the site does not own, while another refuses every request it receives. Routing each site to its best local model reaches 91.8% against 95.4% at no per-call cost. In a live deployment judged by a user, hosting only the two generative sites matches hosting everything (39/43 against 39/43) for 28% of the spend, and the actuation gap the benchmark predicted appears as exactly one case in twenty-six. Benchmark, harness and all records are released at https://github.com/waldiez/slm-callsite-eval.

Large Language Models (LLMs) have seen an impressive wave of advances recently, with models now excelling in a variety of tasks, such as mathematical reasoning and program synthesis. However, their potential to effectively use tools via API calls remains unfulfilled. This is a challenging task even for today's state-of-the-art LLMs such as GPT-4, largely due to their inability to generate accurate input arguments and their tendency to hallucinate the wrong usage of an API call. We release Gorilla, a finetuned LLaMA-based model that surpasses the performance of GPT-4 on writing API calls. When combined with a document retriever, Gorilla demonstrates a strong capability to adapt to test-time document changes, enabling flexible user updates or version changes. It also substantially mitigates the issue of hallucination, commonly encountered when prompting LLMs directly. To evaluate the model's ability, we introduce APIBench, a comprehensive dataset consisting of HuggingFace, TorchHub, and TensorHub APIs. The successful integration of the retrieval system with Gorilla demonstrates the potential for LLMs to use tools more accurately, keep up with frequently updated documentation, and consequently increase the reliability and applicability of their outputs. Gorilla's code, model, data, and demo are available at https://gorilla.cs.berkeley.edu

External tool integration through function-calling is essential for practical language model applications, yet most multilingual models lack reliable tool-use capabilities in non-English languages. Even state-of-the-art multilingual models struggle with determining when to use tools and generating the structured outputs required for function calls, often exhibiting language confusion when prompted in lower-resource languages. This work presents a methodology for adapting existing language models to enable robust tool use in any target language, using Bulgarian as a case study. The approach involves continued training of the BgGPT model series (2.6B, 9B, 27B parameters) on a novel bilingual dataset of 10,035 function-calling examples designed to support standardized protocols like MCP (Model Context Protocol). The research introduces TUCAN (Tool-Using Capable Assistant Navigator), which achieves up to 28.75% improvement in function-calling accuracy over base models while preserving core language understanding, as verified on established Bulgarian benchmarks. Beyond accuracy gains, TUCAN models demonstrate production-ready response formatting with clean, parsable function calls, contrasting with the verbose and inconsistent outputs of base models. The models, evaluation framework, and dataset are released to enable replication for other languages. This work demonstrates a practical approach for extending tool-augmented capabilities beyond English-centric systems.

The ability to use tools is fundamental for large language model (LLM) agents. Given a task, existing systems use LLMs to plan and generate tool calls, which are executed by real-world tools to complete the task. However, tool calls are prone to errors because they are derived merely from LLM intrinsic capabilities. What is more, while it is useful to let LLMs iteratively refine the tool-call sequence using execution results from real tools, this process can be expensive and lead to unsafe results. To improve LLM tool calls and address issues caused by using real tools for refinement, we introduce Gecko, a comprehensive environment that simulates tool responses using a combination of rules and LLMs. Specifically, Gecko checks the validity of tool calls including input arguments and tool names, synthesizes reasonable responses that adhere to the output schema, and assesses whether all task objectives have been achieved. These three types of feedback provided by Gecko allow LLMs to refine their tool calls, forming a simple yet effective test-time scaling method named GATS. On BFCLv3 and τ^2-bench, GATS consistently improves the tool calling performance of various LLMs including GPT-4o, GPT-5, and Gemini-3.0-pro. We further discuss working mechanisms of our method and share future possibilities.

Keyframe selection is a direct way to provide verifiable visual evidence for long-video question answering (QA). Queries differ in what they require, and finding the right frames depends on knowing what to look for. Existing keyframe selectors either score every frame against a single query, or decompose the query into a fixed schema evaluated by a single visual tool. We propose ToolMerge, a keyframe retrieval method based on decomposition and merging: an Large Language Model (LLM) based planner decomposes the query into tool calls and specifies how their per-tool rankings are merged using boolean operators. To evaluate retrieval directly, we construct Molmo-2 Moments (M2M), a benchmark in which every question is anchored to a specific time interval by construction. Across QA, question retrieval, and caption retrieval, ToolMerge is competitive with prior keyframe selectors, most notably on caption retrieval, outperforming other methods by 5%. Code and data can be found at https://github.com/michalsr/ToolMerge .

Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails. We present VeriPhy, an auditable physical-verification system in which a text-only planner compiles the prompt into typed physical obligations and a statically validated execution plan before any frame is observed. During execution, observations gate and scope only declared calls to frozen low-level experts (e.g., segmentation and tracking, counting, eleven typed physical measurements over the resulting tracks, depth, OCR, and audio-event detection). Each action returns a provenance-carrying evidence record whose payload, when usable, is either a typed measurement or an explicitly tagged learned state. Typed resolvers and fixed composition map usable records to a three-valued state (supported, contradicted, or unknown, surfaced as plausible, implausible, or abstain) with full provenance, so that every verdict is traceable to the evidence that produced it. We anchor evaluation in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time. On a 149-clip core carrying 304 such records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims. Recall alone does not separate it from prompting the same backbone monolithically, which reaches 222; what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time and usable as the interface through which a critic verdict could be written back into generation.

Large language models (LLMs) have significantly advanced autonomous agents, particularly in zero-shot tool usage, also known as function calling. This research delves into enhancing the function-calling capabilities of LLMs by exploring different approaches, including prompt formats for integrating function descriptions, blending function-calling and instruction-following data, introducing a novel Decision Token for conditional prompts, leveraging chain-of-thought reasoning, and overcoming multilingual challenges with a translation pipeline. Our key findings and contributions are as follows: (1) Instruction-following data improves both function-calling accuracy and relevance detection. (2) The use of the newly proposed Decision Token, combined with synthetic non-function-call data, enhances relevance detection. (3) A tailored translation pipeline effectively overcomes multilingual limitations, demonstrating significant improvements in Traditional Chinese. These insights highlight the potential for improved function-calling capabilities and multilingual applications in LLMs.

With the advance of large language models (LLMs), the research field of LLM applications becomes more and more popular and the idea of constructing pipelines to accomplish complex tasks by stacking LLM API calls come true. However, this kind of methods face two limitations: narrow information coverage and low fault tolerance. In this work, we propose a novel method called ALLIES. Given an input query, ALLIES leverages LLMs to iteratively generate new queries related to the original query, enabling an iterative reasoning process. By iteratively refining and expanding the scope of the original query, ALLIES captures and utilizes hidden knowledge that may not be directly obtainable through retrieval. We take zero-shot open-domain question answering (ODQA) as an application scene and evaluate ALLIES on the widely-used benchmarks, such as NQ, WebQ and TriviaQA. The experimental results demonstrate that ALLIES significantly outperforms other zero-shot baselines, indicating its effectiveness in tackling those challenges. Our code is available in https://github.com/microsoft/SimXNS/tree/main/ALLIES.

Large language models (LLMs) trained on code completion have been shown to be capable of synthesizing simple Python programs from docstrings [1]. We find that these code-writing LLMs can be re-purposed to write robot policy code, given natural language commands. Specifically, policy code can express functions or feedback loops that process perception outputs (e.g.,from object detectors [2], [3]) and parameterize control primitive APIs. When provided as input several example language commands (formatted as comments) followed by corresponding policy code (via few-shot prompting), LLMs can take in new commands and autonomously re-compose API calls to generate new policy code respectively. By chaining classic logic structures and referencing third-party libraries (e.g., NumPy, Shapely) to perform arithmetic, LLMs used in this way can write robot policies that (i) exhibit spatial-geometric reasoning, (ii) generalize to new instructions, and (iii) prescribe precise values (e.g., velocities) to ambiguous descriptions ("faster") depending on context (i.e., behavioral commonsense). This paper presents code as policies: a robot-centric formulation of language model generated programs (LMPs) that can represent reactive policies (e.g., impedance controllers), as well as waypoint-based policies (vision-based pick and place, trajectory-based control), demonstrated across multiple real robot platforms. Central to our approach is prompting hierarchical code-gen (recursively defining undefined functions), which can write more complex code and also improves state-of-the-art to solve 39.8% of problems on the HumanEval [1] benchmark. Code and videos are available at https://code-as-policies.github.io

Large language models (LLMs) with billions of parameters have sparked a new wave of exciting AI applications. However, their high computational costs and memory demands during inference pose significant challenges. Adaptive sparse activation inference, which activates only a small number of neurons for each token, offers a novel way to accelerate model inference without degrading performance, showing great potential for resource-constrained hardware devices. Nevertheless, existing methods predict activated neurons based on individual tokens with additional MLP, which involve frequent changes in activation maps and resource calls, limiting the acceleration benefits of sparse activation. In this paper, we introduce CoreInfer, an MLP-free adaptive sparse activation inference method based on sentence-level prediction. Specifically, we propose the concept of sentence-wise core neurons, which refers to the subset of neurons most critical for a given sentence, and empirically demonstrate its effectiveness. To determine the core neurons, we explore the correlation between core neurons and the sentence's semantics. Remarkably, we discovered that core neurons exhibit both stability and similarity in relation to the sentence's semantics -- an insight overlooked by previous studies. Building on this finding, we further design two semantic-based methods for predicting core neurons to fit different input scenarios. In CoreInfer, the core neurons are determined during the pre-filling stage and fixed during the encoding stage, enabling zero-cost sparse inference. We evaluated the model generalization and task generalization of CoreInfer across various models and tasks. Notably, on an NVIDIA TITAN XP GPU, CoreInfer achieved a 10.33 times and 2.72 times speedup compared to the Huggingface implementation and PowerInfer, respectively.

Restless multi-armed bandits (RMAB) have demonstrated success in optimizing resource allocation for large beneficiary populations in public health settings. Unfortunately, RMAB models lack flexibility to adapt to evolving public health policy priorities. Concurrently, Large Language Models (LLMs) have emerged as adept automated planners across domains of robotic control and navigation. In this paper, we propose a Decision Language Model (DLM) for RMABs, enabling dynamic fine-tuning of RMAB policies in public health settings using human-language commands. We propose using LLMs as automated planners to (1) interpret human policy preference prompts, (2) propose reward functions as code for a multi-agent RMAB environment, and (3) iterate on the generated reward functions using feedback from grounded RMAB simulations. We illustrate the application of DLM in collaboration with ARMMAN, an India-based non-profit promoting preventative care for pregnant mothers, that currently relies on RMAB policies to optimally allocate health worker calls to low-resource populations. We conduct a technology demonstration in simulation using the Gemini Pro model, showing DLM can dynamically shape policy outcomes using only human prompts as input.