Daily Papers

Tool-using agents are increasingly expected to operate across realistic professional workflows, where they must interpret multimodal inputs, coordinate external tools, inspect intermediate artifacts, and revise their actions before producing a final result. Existing benchmarks, however, often evaluate tool use, computer use, and multimodal reasoning in isolation, leaving a gap between benchmark settings and end-to-end omni-modal tool use in the real world. To address this gap, we introduce MM-ToolBench, a benchmark and evaluation harness for task-oriented omni-modal tool use. MM-ToolBench contains 100 executable tasks from two macro task families, Customer Service and Intelligent Creation, covering 20 subcategory slices and supported by 27 MCP servers with 324 tools. The central design of MM-ToolBench is closed-loop multimodal verification: agents must execute tools, inspect rendered or transformed artifacts, and self-correct when outputs fail task-specific requirements. To make such evaluation scalable and verifiable, MM-ToolBench couples MCP-based execution with task-specific grounded evaluators and a semi-automated construction pipeline for scenario discovery, task instantiation, evaluator synthesis, and human audit. Experiments on 15 contemporary agentic models show that MM-ToolBench remains highly challenging: Claude Opus 4.6, commonly regarded as one of the strongest coding-agent models, achieves only 32.0% task success, far below the 94.0% human benchmark. We envision MM-ToolBench as a practical foundation for evaluating and advancing next-generation omni-modal tool-using agents through closed-loop multimodal verification.

Tool-using agents often incur substantial computational cost due to long reasoning chains and iterative tool usage. In practical scenarios, many tasks become infeasible under constrained tool environments, where the capabilities required for successful task completion are unavailable. Detecting infeasible tasks and stopping execution early can significantly reduce unnecessary execution cost. In this work, we propose FeasiGen, an automatic pipeline for constructing infeasible agent tasks by identifying the critical tools required for successful task completion. Our approach extracts tool-calling traces from successful executions across multiple agent systems, identifies critical tools consistently shared across diverse execution strategies, and masks these tools to automatically transform solvable tasks into infeasible ones. Human verification confirms that the infeasibility annotations for our constructed tasks achieve over 94% accuracy. We further introduce feasibility-aware evaluation metrics for measuring whether agents can recognize infeasible tasks and stop execution appropriately. Extensive evaluations across nine models reveal substantially weak infeasibility detection ability, with false continue rate reaching up to 73.9%. We further observe that multi-agent architectures significantly reduce erroneous execution under infeasible conditions.

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

Tool-using LLM agents spend wall-clock time not only on model inference but also in serial action--observation turns, where each tool call, environment transition, and observation can delay subsequent decisions. We introduce Speculative Macro Commit (SMC), a runtime mechanism for a two-tier agent system: a large authoritative actor model produces the official trajectory, while a faster speculative drafter model continuously predicts and executes future action chains on an isolated environment snapshot. SMC mines recurring multi-action skeletons from training traces and stores them in a macro library used to match against action chains predicted by the drafter at runtime. When the actor's next tool call matches the first drafted action, SMC commits the remaining pre-executed draft steps, together with their observations, to the official trajectory. Using Qwen3.5-27B INT4 as the authoritative actor model and Qwen3.5-4B as the speculative drafter model, SMC matches the sequential agent's overall accuracy while reducing latency by 10.23\% over the Speculative Actions (SA) baseline and 18.59\% over sequential execution on the τ^2-Bench Telecom subset. On AppWorld, SMC reduces wall time by 7.7\% over SA baseline and 44.9\% over sequential execution, with a small reduction in task completion. Overall, SMC provides a practical way to reuse multi-step speculative execution and reduce agent latency beyond single-step speculative actions. Our code is publicly available https://github.com/zeyuliu1037/speculative-macro-commit{magenta{here}}.

We present Test-Driven AI Agent Definition (TDAD), a methodology that treats agent prompts as compiled artifacts: engineers provide behavioral specifications, a coding agent converts them into executable tests, and a second coding agent iteratively refines the prompt until tests pass. Deploying tool-using LLM agents in production requires measurable behavioral compliance that current development practices cannot provide. Small prompt changes cause silent regressions, tool misuse goes undetected, and policy violations emerge only after deployment. To mitigate specification gaming, TDAD introduces three mechanisms: (1) visible/hidden test splits that withhold evaluation tests during compilation, (2) semantic mutation testing via a post-compilation agent that generates plausible faulty prompt variants, with the harness measuring whether the test suite detects them, and (3) spec evolution scenarios that quantify regression safety when requirements change. We evaluate TDAD on SpecSuite-Core, a benchmark of four deeply-specified agents spanning policy compliance, grounded analytics, runbook adherence, and deterministic enforcement. Across 24 independent trials, TDAD achieves 92% v1 compilation success with 97% mean hidden pass rate; evolved specifications compile at 58%, with most failed runs passing all visible tests except 1-2, and show 86-100% mutation scores, 78% v2 hidden pass rate, and 97% regression safety scores. The implementation is available as an open benchmark at https://github.com/f-labs-io/tdad-paper-code.

REST APIs play important roles in enriching the action space of web agents, yet most API-based agents rely on curated and uniform toolsets that do not reflect the complexity of real-world APIs. Building tool-using agents for arbitrary domains remains a major challenge, as it requires reading unstructured API documentation, testing APIs and inferring correct parameters. We propose Doc2Agent, a scalable pipeline to build agents that can call Python-based tools generated from API documentation. Doc2Agent generates executable tools from API documentations and iteratively refines them using a code agent. We evaluate our approach on real-world APIs, WebArena APIs, and research APIs, producing validated tools. We achieved a 55\% relative performance improvement with 90\% lower cost compared to direct API calling on WebArena benchmark. A domain-specific agent built for glycomaterial science further demonstrates the pipeline's adaptability to complex, knowledge-rich tasks. Doc2Agent offers a generalizable solution for building tool agents from unstructured API documentation at scale.

Interactive tool-using agents must solve real-world tasks via multi-turn interaction with both humans and external environments, requiring dialogue state tracking, multi-step tool execution, while following complex instructions. Post-training such agents is challenging because synthesis for high-quality multi-turn tool-use data is difficult to scale, and reinforcement learning (RL) could face noisy signals caused by user simulation, leading to degraded training efficiency. We propose a unified framework that combines a self-evolving data agent with verifier-based RL. Our system, EigenData, is a hierarchical multi-agent engine that synthesizes tool-grounded dialogues together with executable per-instance checkers, and improves generation reliability via closed-loop self-evolving process that updates prompts and workflow. Building on the synthetic data, we develop an RL recipe that first fine-tunes the user model and then applies GRPO-style training with trajectory-level group-relative advantages and dynamic filtering, yielding consistent improvements beyond SFT. Evaluated on tau^2-bench, our best model reaches 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom, matching or exceeding frontier models. Overall, our results suggest a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation.

Large language models are increasingly deployed as agents that solve tasks by interacting with external tool environments. Although recent tool-use benchmarks increasingly cover complex task settings, they still largely assume clean, stable, and trustworthy tool environments, leaving tool-environment unreliability insufficiently examined. We introduce ToolBench-X, a benchmark for evaluating agents under recoverable reliability hazards. ToolBench-X contains executable multi-step tasks across diverse domains and sequential, parallel, and mixed workflows, each paired with deterministic tools and a canonical final answer for automatic evaluation. Starting from clean tool environments, ToolBench-X injects five structured hazard types: Specification Drift, Invocation Error, Execution Failure, Output Drift, and Cross-source Conflict. Crucially, each injected instance remains solvable through at least one valid recovery path, such as retrying, fallback, verification, or cross-checking. Experiments reveal a substantial reliability gap: agents that perform well with reliable tools often fail under recoverable hazards. Further analysis shows that failures are driven less by tool-use volume or inference budget than by limited hazard diagnosis and ineffective recovery. Targeted recovery hints recover many failed tasks, while test-time scaling yields more limited gains. These results suggest that tool-use evaluation should move beyond function-call accuracy toward task completion under unreliable tool environments. The code and data is available at https://github.com/Foreverskyou/ToolBench-X.

Recent video editing models have converged on a unified conditioning design: a single diffusion transformer jointly consumes text, source video, and reference images, and one set of weights covers replacement, removal, style transfer, and reference-driven insertion. The design is flexible, but it assumes that the user already provides model-ready text, reference images, and spatial grounding for local edits, which real requests often omit. We present Aurora, an agentic video editing framework that pairs a tool-augmented vision-language model (VLM) agent with a unified video diffusion transformer. The VLM agent maps a raw user request to a structured edit plan aligned with the transformer's conditioning channels, thereby resolving textual and visual underspecification before generation. We train the VLM agent with supervised data for complete edit planning and reference-image selection, together with preference pairs for robust tool use and instruction refinement. We introduce AgentEdit-Bench to evaluate agent-enhanced video editing under textual and visual underspecification. Experiments on AgentEdit-Bench and two existing video editing benchmarks show that Aurora improves over instruction-only baselines and that the VLM agent transfers to compatible frozen video editing models. Project page: https://yeates.github.io/Aurora-Page

Large language model agents solve tasks through long sequences of tool calls, and the prevailing design has a single strong model generate every step, although many steps are routine enough for a much smaller model. Existing ways of sharing this work keep the strong model as the generator: routers and collaborative agents decide which model handles each query or step, and speculative methods check a small model's drafts one step at a time, usually against the strong model's own action. We introduce SpecAgent, a collaboration protocol that turns the strong model from the author of each step into a reviewer of steps drafted by a weak model. The weak model executes read-only tool calls as it drafts, so that every step is grounded in real observations, and holds the first action that would modify the environment. The strong model then verifies the drafted segment in a single prefill pass, commits the longest prefix it judges semantically valid, and repairs the first rejected step. Because verification replaces autoregressive decoding with parallel prefill, the strong model decodes only short verdicts and occasional repairs rather than every step. On τ^2-Bench, SpecAgent comes within one point of the strong model's success rate while roughly halving memory traffic and wall time, and on AutomationBench and VitaBench it recovers 44-75% of the quality gap while lowering both. In an edge-cloud deployment on MCP-Atlas with three cloud models, it closes 53-75% of the weak-strong quality gap while the strong model generates at most half as many tokens. These results suggest that a strong model need not author every step of an agent: reviewing grounded drafts retains most of its quality while cutting both latency and memory traffic.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

While Large Language Models (LLMs) have evolved into tool-using agents, they remain brittle in long-horizon interactions. Unlike mathematical reasoning where errors are often rectifiable via backtracking, tool-use failures frequently induce irreversible side effects, making accurate step-level verification critical. However, existing process-level benchmarks are predominantly confined to closed-world mathematical domains, failing to capture the dynamic and open-ended nature of tool execution. To bridge this gap, we introduce AgentProcessBench, the first benchmark dedicated to evaluating step-level effectiveness in realistic, tool-augmented trajectories. The benchmark comprises 1,000 diverse trajectories and 8,509 human-labeled step annotations with 89.1% inter-annotator agreement. It features a ternary labeling scheme to capture exploration and an error propagation rule to reduce labeling ambiguity. Extensive experiments reveal key insights: (1) weaker policy models exhibit inflated ratios of correct steps due to early termination; (2) distinguishing neutral and erroneous actions remains a significant challenge for current models; and (3) process-derived signals provide complementary value to outcome supervision, significantly enhancing test-time scaling. We hope AgentProcessBench can foster future research in reward models and pave the way toward general agents. The code and data are available at https://github.com/RUCBM/AgentProcessBench.

As AI agents increasingly operate in open, real-world environments, they require a deep synergy of multimodal perception, tool invocation with multi-hop reasoning, and dynamic interaction with users. However, existing benchmarks fail to jointly evaluate these capabilities due to challenges in designing strictly coupled multi-capability tasks, simulating natural and task-constrained user feedback, and ensuring objective evaluation of dynamic interaction. To bridge this gap, we introduce EgoBench, the first interactive multimodal benchmark for tool-using agents. EgoBench comprises 1,045 egocentric-video-grounded tasks covering four daily scenarios, along with a user-agent-tool interactive environment for evaluation. We implement a three-stage synergistic pipeline through which each task is designed to enforce the joint application of visual perception and tool-augmented multi-hop reasoning. We additionally develop a multi-agent simulated user within EgoBench to evaluate agents' interaction capabilities, which generates high-fidelity, task-aligned responses to agents. Furthermore, we establish a deterministic joint validation framework that guarantees objective assessment through process-based and result-based equivalence. Benchmarking eight SOTA video-MLLM agents on EgoBench reveals a severe performance ceiling: the best model achieves only 30.62% accuracy in the best-performing scenario, averaging 19.43% across all four scenarios. Finally, we conduct a multi-dimensional error analysis to disentangle failure modes, exposing capability bottlenecks for advancing future AI agents.

Tool-using LLM agents still struggle in open-world settings with large tool pools, long-horizon objectives, wild constraints, and unreliable tool states. For scalable and realistic training and testing, we introduce an open-world tool-using environment, built on 5,571 format unified tools across 204 commonly used apps. It includes a task creation engine that synthesizes long-horizon, multi-tool workflows with wild constraints, and a state controller that injects interruptions and failures to stress-test robustness. On top of this environment, we develop a tool select-then-execute agent framework with a planner-actor decomposition to separate deliberate reasoning and self-correction from step-wise execution. Comprehensive evaluation of state-of-the-art LLMs reveals the misalignment between tool planning and execution abilities, the constraint following weakness of existing LLMs, and DeepSeek-v3.2's strongest robustness. Finally, we collect 1,170 trajectories from our environment to fine-tune LLMs, achieving superior performance to baselines using 119k samples, indicating the environment's value as both a realistic benchmark and a data engine for tool-using agents. Our code and data will be publicly released.

An agent whose tool keeps returning nothing useful should stop relying on it. In a retrieval environment with controlled source failures, we separate how agents judge results from what they do. We compare stopping at the same step after longer and shorter runs of results the agent judged useless; this contrast is zero for clock- or deadline-driven stopping. Where we record their judgments, the seven agents we test call a failing source's results useless 97-100% of the time, yet most of them rarely stop on that judgment. Prompt cues change when they stop but not what they stop on. Permission to answer from memory and a reasoning mode can bring early stops regardless of evidence, a stated budget moves the 7-8B models' stops to the deadline, and a stopping rule or call cost in the prompt is followed at most partly. Stopping follows the evidence only when the harness enforces an integration step that makes the agent answer after five consecutive results it judged useless. This step raises failing-source success for every model, keeps the stopping point fixed when the budget doubles, and needs no extra judgment call when the agent states its judgments. A pre-registered replication on 300 fresh questions confirms the dissociation and the rule's effect.

Reinforcement Learning (RL) is essential for evolving Large Language Models (LLMs) into autonomous agents capable of long-horizon planning, yet a practical recipe for scaling RL in complex, multi-turn environments remains elusive. This paper presents a systematic empirical study using TravelPlanner, a challenging testbed requiring tool orchestration to satisfy multifaceted constraints. We decompose the agentic RL design space along 5 axes: reward shaping, model scaling, data composition, algorithm selection, and environmental stability. Our controlled experiments yield 7 key takeaways, e.g., (1) reward and algorithm choices are scale-dependent as smaller models benefit from staged rewards and enhanced exploration, whereas larger models converge efficiently with simpler dense rewards, (2) ~ 1K training samples with a balanced difficulty mixture mark a sweet spot for both in-domain and out-of-domain performance, and (3) environmental stability is critical to prevent policy degradation. Based on our distilled recipe, our RL-trained models achieve state-of-the-art performance on TravelPlanner, significantly outperforming leading LLMs.

LLM agents struggle with regulatory audit replay: when asked to reproduce a flagged transaction decision with identical inputs, many deployments fail to return consistent results. We introduce the Determinism-Faithfulness Assurance Harness (DFAH), a framework for measuring trajectory determinism, decision determinism, and evidence-conditioned faithfulness in tool-using agents deployed in financial services. Across 4,700+ agentic runs (7 models, 4 providers, 3 financial benchmarks with 50 cases each at T=0.0), we find that decision determinism and task accuracy are not detectably correlated (r = -0.11, 95% CI [-0.49, 0.31], p = 0.63, n = 21 configurations): models can be deterministic without being accurate, and accurate without being deterministic. Because neither metric predicts the other in our sample, both must be measured independently, which is precisely what DFAH provides. Small models (7-20B) achieve near-perfect determinism through rigid pattern matching at the cost of accuracy (20-42%), while frontier models show moderate determinism (50-96%) with variable accuracy. No model achieves both perfect determinism and high accuracy, supporting DFAH's multi-dimensional measurement approach. We provide three financial benchmarks (compliance triage, portfolio constraints, and DataOps exceptions; 50 cases each) together with an open-source stress-test harness. Across these benchmarks and DFAH evaluation settings, Tier 1 models with schema-first architectures achieved determinism levels consistent with audit replay requirements.

Language-model agents increasingly use tools to act on external systems. Earlier actions can alter files, permissions, database records, or other state, making a later routine-looking action harmful. Yet the visible interaction may not reveal the underlying state needed to assess that action. We formulate attack and defense as partially observed state control in SEAD, deriving their design requirements from this shared execution process. Because attackers supply instructions while the target chooses concrete actions, DART decomposes harmful goals into locally plausible steps and uses feedback from actual tool execution to guide trajectory search. The defender must decide before execution with incomplete state evidence. SAGE can therefore investigate relevant state through read-only queries before allowing or blocking each action, including those proposed after a block. We construct an environment-verifiable dataset integrating controlled initial states, replayable tool environments, and task-specific executable checks. Across four target models, DART improves semantic attack success by 18.8--35.9 percentage points over the competing baseline, with consistent gains under executable verification. On recorded trajectories, SAGE preserves 95.79% of benign trajectories while intercepting 92.73% of harmful paths by the harm-enabling boundary. In online attack-defense evaluation, it reduces DART's executable attack success from 48.0% to 4.0%. SAGE remains effective across four attack methods and generalizes to out-of-domain environments. Our code and data is available at https://github.com/EverywhereSafety/SEAD.

Language model agents are increasingly effective in solving realistic tasks through multi-turn tool use. However, training reliable tool-using agents remains challenging in practice. While reinforcement learning provides an on-policy paradigm for improving agents from their own environment interactions, its effectiveness depends heavily on the training task distribution. When tasks are fixed before training, the task distribution can become increasingly mismatched with the policy's evolving capabilities, causing many rollouts to be spent on uninformative tasks. We propose SENTINEL, a failure-driven reinforcement learning framework that turns the Solver's rollout failures into targeted training tasks. SENTINEL follows a Controller--Proposer--Solver loop: the Controller analyzes failed trajectories and summarizes recurring error patterns, the Proposer generates executable tasks that stress these weaknesses, and the Solver is trained on the targeted tasks. On Tau2-Bench Retail with Qwen3-4B-Thinking-2507, SENTINEL improves Pass\^{}1 from 66.4 to 74.9 and outperforms RL on general synthetic tasks across Pass\^{}k metrics. These results demonstrate that model failures provide an effective and scalable source of targeted training signal for improving tool-using language model agents.

Reward-guided search methods have demonstrated strong potential in enhancing tool-using agents by effectively guiding sampling and exploration over complex action spaces. As a core design, those search methods utilize process reward models (PRMs) to provide step-level rewards, enabling more fine-grained monitoring. However, there is a lack of systematic and reliable evaluation benchmarks for PRMs in tool-using settings. In this paper, we introduce ToolPRMBench, a large-scale benchmark specifically designed to evaluate PRMs for tool-using agents. ToolPRMBench is built on top of several representative tool-using benchmarks and converts agent trajectories into step-level test cases. Each case contains the interaction history, a correct action, a plausible but incorrect alternative, and relevant tool metadata. We respectively utilize offline sampling to isolate local single-step errors and online sampling to capture realistic multi-step failures from full agent rollouts. A multi-LLM verification pipeline is proposed to reduce label noise and ensure data quality. We conduct extensive experiments across large language models, general PRMs, and tool-specialized PRMs on ToolPRMBench. The results reveal clear differences in PRM effectiveness and highlight the potential of specialized PRMs for tool-using. Code and data will be released at https://github.com/David-Li0406/ToolPRMBench.

Tool-using agents based on Large Language Models (LLMs) excel in tasks such as mathematical reasoning and multi-hop question answering. However, in long trajectories, agents often trigger excessive and low-quality tool calls, increasing latency and degrading inference performance, making managing tool-use behavior challenging. In this work, we conduct entropy-based pilot experiments and observe a strong positive correlation between entropy reduction and high-quality tool calls. Building on this finding, we propose using entropy reduction as a supervisory signal and design two reward strategies to address the differing needs of optimizing tool-use behavior. Sparse outcome rewards provide coarse, trajectory-level guidance to improve efficiency, while dense process rewards offer fine-grained supervision to enhance performance. Experiments across diverse domains show that both reward designs improve tool-use behavior: the former reduces tool calls by 72.07% compared to the average of baselines, while the latter improves performance by 22.27%. These results position entropy reduction as a key mechanism for enhancing tool-use behavior, enabling agents to be more adaptive in real-world applications.

Long-horizon tool-using agents must reason over user goals, domain policies, tool calls, simulator state, and delayed verifiable rewards. Reinforcement learning (RL) is a natural fit for this setting, but multi-turn on-policy rollouts create long contexts, while model-specific attention layers may require custom masks and learned sink normalization. We present SINKFLEX-RL, a modular training system for RL in dual-control tool-use environments. The system combines a Gymnasium-compatible environment wrapper, a VERL-style rollout dataflow, group-relative policy optimization without a separate value model, and a sink-aware FlexAttention path designed to preserve model-specific sink scaling under causal and sliding-window masks. In a preliminary Tau2Bench retail run, validation reward (mean@1) rises from 0.25 early in training to 0.44 later in the observed training window, while training-score and trajectory-reward proxies also trend upward. In a fixed-configuration memory benchmark, the optimized attention path reduces peak VRAM from 28.06GB to 22.52GB at 4096 tokens, a 19.7% reduction, and runs the measured 8192-token configuration using 25.53~GB where the eager baseline runs out of memory. These results illustrate the value of integrating environment interfaces, RL dataflow, and attention-kernel design for memory-feasible long-horizon agent training.

As AI agents increasingly act through external tools, tool results are no longer simple outputs but execution artifacts that must remain correctly associated with the context that produced them. Yet as results persist across caches, recovery mechanisms, and shared runtime state, this association can break, causing a real and unmodified result to be consumed in an incompatible execution context and lead the agent to an unsafe action. We characterize this failure as tool-result misbinding. To systematically study this failure, we introduce ReplayBench, a benchmark that uses controlled tool-result replay to evaluate misbinding across 19 representative agent scenarios, 4 failure properties, and 3 replay mechanisms. Across eight LLMs, three agent frameworks, and diverse stateful workflows, we find widespread replay-induced failures, with a mean Final-State Error Rate of 0.649 under cross-task replay. Stage-wise analysis localizes these failures to model-side action generation. Further analyses show that replay susceptibility varies with task similarity, shows little dependence on object similarity, and remains high under result staleness and implicit environment changes. These findings motivate runtime validity mechanisms that eliminate final-state errors in our evaluated settings while preserving legitimate recovery. The results call for explicit contextual binding in tool-using agent runtimes. Our work is public at https://github.com/Phitooo/Replaybench.

Ultra-long egocentric video understanding requires reasoning over temporally sparse evidence distributed across hours or days, challenging current multimodal models with limited context and the grounding of key video segments. While Chain-of-Tool-Thought (CoTT) agent systems enable iterative retrieval and inspection, they suffer from error propagation due to rigid zoom-in strategies that lack recovery mechanisms. In this work, we address these challenges through SCOUT (Self-Checking Chain-Of-Tool-thought), a recovery-aware agentic framework introducing an adaptive policy that evaluates intermediate tool observations and dynamically trades off exploitation (zoom-in) and exploration (region switching), enabling robust multi-hop reasoning over extremely long horizons. However, training such multi-turn tool-using agents remains challenging, as existing RL methods rely on sparse outcome-level rewards and lack supervision over extended decision trajectories, resulting in suboptimal credit assignment for long-horizon reasoning. To address this, we develop UPS-GRPO, an uncertainty-prioritized policy optimization method that concentrates exploration on high-uncertainty post-tool states while preserving sample efficiency. We further introduce a turn-level advantage decomposition that integrates outcome rewards with tool-grounded temporal alignment rewards for improved credit assignment. Experiments show that SCOUT achieves state-of-the-art results on ultra-long egocentric benchmarks, while remaining competitive on shorter-horizon long-video settings.

Training reliable tool-augmented agents remains a significant challenge, largely due to the difficulty of credit assignment in multi-step reasoning. While process-level reward models offer a promising direction, existing LLM-based judges often produce noisy and inconsistent signals because they lack fine-grained, task-specific rubrics to distinguish high-level planning from low-level execution. In this work, we introduce SCRIBE (Skill-Conditioned Reward with Intermediate Behavioral Evaluation), a reinforcement learning framework that intervenes at a novel mid-level abstraction. SCRIBE grounds reward modeling in a curated library of skill prototypes, transforming open-ended LLM evaluation into a constrained verification problem. By routing each subgoal to a corresponding prototype, the reward model is equipped with precise, structured rubrics that substantially reduce reward variance. Experimental results show that SCRIBE achieves state-of-the-art performance across a range of reasoning and tool-use benchmarks. In particular, it improves the AIME25 accuracy of a Qwen3-4B model from 43.3% to 63.3%, and significantly increases success rates in complex multi-turn tool interactions. Further analysis of training dynamics reveals a co-evolution across abstraction levels, where mastery of mid-level skills consistently precedes the emergence of effective high-level planning behaviors. Finally, we demonstrate that SCRIBE is additive to low-level tool optimizations, providing a scalable and complementary pathway toward more autonomous and reliable tool-using agents.

AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn's intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.

Tool-using agent systems powered by large language models (LLMs) are increasingly deployed across web, app, operating-system, and transactional environments. Yet existing safety benchmarks still emphasize explicit risks, potentially overstating a model's ability to judge deceptive or ambiguous trajectories. To address this gap, we introduce ROME (Red-team Orchestrated Multi-agent Evolution), a controlled benchmark-construction pipeline that rewrites known unsafe trajectories into more deceptive evaluation instances while preserving their underlying risk labels. Starting from 100 unsafe source trajectories, ROME produces 300 challenge instances spanning contextual ambiguity, implicit risks, and shortcut decision-making. Experiments show that these challenge sets substantially degrade safety-judgment performance, with hidden-risk cases remaining particularly non-trivial even for recent frontier models. We further study ARISE (Analogical Reasoning for Inference-time Safety Enhancement), a retrieval-guided inference-time enhancement that retrieves ReAct-style analogical safety trajectories from an external analogical base and injects them as structured reasoning exemplars. ARISE improves judgment quality without retraining, but is best viewed as a task-specific robustness enhancement rather than a standalone safety guarantee. Together, ROME and ARISE provide practical tools for stress-testing and improving agent safety judgment under deceptive distribution shifts.

Tool-using agents that act in the world need to be both useful and safe. Well-calibrated model confidences can be used to weigh the risk versus reward of potential actions, but prior work shows that many models are poorly calibrated. Inspired by interpretability literature exploring the internals of models, we propose a novel class of model-internal confidence estimators (MICE) to better assess confidence when calling tools. MICE first decodes from each intermediate layer of the language model using logitLens and then computes similarity scores between each layer's generation and the final output. These features are fed into a learned probabilistic classifier to assess confidence in the decoded output. On the simulated trial and error (STE) tool-calling dataset using Llama3 models, we find that MICE beats or matches the baselines on smoothed expected calibration error. Using MICE confidences to determine whether to call a tool significantly improves over strong baselines on a new metric, expected tool-calling utility. Further experiments show that MICE is sample-efficient, can generalize zero-shot to unseen APIs, and results in higher tool-calling utility in scenarios with varying risk levels. Our code is open source, available at https://github.com/microsoft/mice_for_cats.

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

AI agents are increasingly transacting on behalf of users -- delegating tasks, spending budgets, and negotiating with unfamiliar counterparties. Unlike human marketplaces, which operate under institutional designs refined over centuries, the rules governing emerging agent marketplaces are being built ad-hoc, and early choices tend to lock in. Understanding what dynamics these rules produce is urgent. We present diagon, a programmable market system serving as a rule-agnostic experimental testbed for institutional design in emerging agent cognitive-labour markets. diagon makes institutional choices experimentally manipulable: heterogeneous tool-using agents post jobs, bid, negotiate, execute, pay, and accumulate reputation, with every mechanism end-to-end observable. We instantiate one market form to demonstrate diagon. We find that market exchange generates more productivity gains over self-sufficient agents, but these gains depend strongly on institutional structure; for example, interventions such as identity transparency and stronger competitive selection can degrade market performance rather than improve it. These findings highlight concrete design requirements for the economic infrastructure of the agent era. Code and data are available at https://github.com/assassin808/diagon.

Tool-using large language model (LLM) agents produce long, multi-turn trajectories, making gradient-based post-training memory-intensive. Evolution strategies (ES) enable memory-efficient full-parameter post-training without backpropagation and can eventually match the performance of gradient-based reinforcement learning (RL). However, resource-constrained settings typically offer only a few GPUs, so the high GPU-hour requirements of ES translate into prohibitively long training times. To address this, we introduce Cooperative Parameter-subspace Evolution Strategy (CoPES), a cooperative coevolutionary method that decomposes the full parameter space into lower-dimensional subspaces and searches over them cooperatively to improve optimization efficiency. We post-train a Qwen3.5-4B tool-using agent for the math task and evaluate it on five benchmarks of varying difficulty. Under the GPU-hour budget of full-parameter GRPO's best validation checkpoint, CoPES recovers 92% of GRPO's validation-accuracy gain, versus 67% for standard ES, while its theoretical GPU memory requirement is less than one-eighth that of full-parameter GRPO. It consistently outperforms standard ES and LoRA-based GRPO on all evaluated pass@k metrics across the five benchmarks. Additional experiments further show the advantage of CoPES on the question-answering task. These results demonstrate an improved trade-off between memory requirements and training time for agentic LLM post-training under resource constraints. The code is open-sourced in https://github.com/MetaronWang/CoPES

Agent skills are Markdown documents that a tool-using agent reads as instructions and then acts on with the privileges of the session that loaded them. Installing one grants authority (the standing to direct the agent, not new privilege), and the size of that grant is what we measure. We present a two-phase attack-surface profiling methodology that combines whole-bundle static analysis and dependency provenance checking with instrumented sandbox execution and separate model-based assessment of the resulting traces. We apply the methodology to 6,309 listings from two marketplaces, selecting listings for execution without conditioning that decision on a prior threat or risk signal. Half the corpus carries evidence of shell-execution capability; the corresponding shares are 35% for filesystem read, 31% each for filesystem write and network egress, and 13% for credential access; seven listings in ten carry at least one elevated capability. That is the surface a user accepts at install time, and almost none of it is visible then. A permission model exists but is close to inert. Authors may list the tools a skill pre-approves in its own frontmatter, which removes confirmation prompts rather than restricting what the skill can reach; of the 6,061 listings whose frontmatter parses, 770 declare anything at all, the vocabulary cannot express credential or environment access, and nothing checks what it does express against the bundle. Declaring is also not declaring accurately: 59% of declaring listings understate what their bundle can do. ClawHub declares roughly twice as often as skillsmp, yet under-declares on 88% of its declarations against 31%. Triage at the marketplace boundary therefore cannot rest on what a listing declares. It must rest on what the bundle can do. These results position capability-oriented attack-surface profiling as a necessary complement to maliciousness detection.

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

A multi-agent AI system (MAS) is composed of multiple autonomous agents that interact, exchange information, and make decisions based on internal generative models. Recent advances in large language models and tool-using agents have made MAS increasingly practical in areas like scientific discovery and collaborative automation. However, key questions remain: When are MAS more effective than single-agent systems? What new safety risks arise from agent interactions? And how should we evaluate their reliability and structure? This paper outlines a formal framework for analyzing MAS, focusing on two core aspects: effectiveness and safety. We explore whether MAS truly improve robustness, adaptability, and performance, or merely repackage known techniques like ensemble learning. We also study how inter-agent dynamics may amplify or suppress system vulnerabilities. While MAS are relatively new to the signal processing community, we envision them as a powerful abstraction that extends classical tools like distributed estimation and sensor fusion to higher-level, policy-driven inference. Through experiments on data science automation, we highlight the potential of MAS to reshape how signal processing systems are designed and trusted.

Diffusion large language models (DLLMs) have emerged as an alternative to autoregressive (AR) decoding with appealing efficiency and modeling properties, yet their implications for agentic multi-step decision making remain underexplored. We ask a concrete question: when the generation paradigm is changed but the agent framework and supervision are held fixed, do diffusion backbones induce systematically different planning and tool-use behaviors, and do these differences translate into end-to-end efficiency gains? We study this in a controlled setting by instantiating DLLM and AR backbones within the same agent workflow (DeepDiver) and performing matched agent-oriented fine-tuning on the same trajectory data, yielding diffusion-backed DLLM Agents and directly comparable AR agents. Across benchmarks and case studies, we find that, at comparable accuracy, DLLM Agents are on average over 30% faster end to end than AR agents, with some cases exceeding 8x speedup. Conditioned on correct task completion, DLLM Agents also require fewer interaction rounds and tool invocations, consistent with higher planner hit rates that converge earlier to a correct action path with less backtracking. We further identify two practical considerations for deploying diffusion backbones in tool-using agents. First, naive DLLM policies are more prone to structured tool-call failures, necessitating stronger tool-call-specific training to emit valid schemas and arguments. Second, for multi-turn inputs interleaving context and action spans, diffusion-style span corruption requires aligned attention masking to avoid spurious context-action information flow; without such alignment, performance degrades. Finally, we analyze attention dynamics across workflow stages and observe paradigm-specific coordination patterns, suggesting stronger global planning signals in diffusion-backed agents.

Tool-calling benchmarks are increasingly used to rank language-model agents, yet their scores are often treated as ground truth without validating the evaluators themselves. We present a systematic validity and reproducibility audit of four major tool-calling benchmark families: BFCL v4, τ2-Bench, LiveMCPBench, and MCP-Atlas. Across 496 expert-reviewed benchmark tasks, we find 92 evaluator-human disagreements, corresponding to an 18.5% misalignment rate. The failures are not isolated annotation mistakes: deterministic benchmarks exhibit brittle state matching, trajectory lock-in, incorrect ground truths, substring-based communication failures, and reward-basis misalignment, while LLM-judge benchmarks exhibit rubric drift, hallucinated completion, answer-only scoring, and substantial run-to-run variance. In LiveMCPBench, 23 repeated evaluations of the same setup produce scores ranging from 57.9% to 76.8%, a spread of 18.9 percentage points, large enough to change leaderboard conclusions. These results show that current tool-calling scores can reflect evaluator artifacts rather than agent capability. We introduce a unified taxonomy of tool-calling evaluation failures, release trace-level audit artifacts and corrected evaluation components, and argue for decomposed metrics that separately measure tool invocation, task completion, and outcome verification. Our findings suggest that progress in tool-using agents requires benchmarks whose evaluators are themselves reproducible, auditable, and aligned with human judgments of task success. We further introduce Tool-Veritas, a configurable benchmark that combines deterministic state verification with optional qualitative judging, and Harness Lab, an open-source system for benchmark execution, trace inspection, repeated-run comparison, and evaluator debugging.

Data agents are an emerging paradigm that leverages large language models (LLMs) and tool-using agents to automate data management, preparation, and analysis tasks. However, the term "data agent" is currently used inconsistently, conflating simple query responsive assistants with aspirational fully autonomous "data scientists". This ambiguity blurs capability boundaries and accountability, making it difficult for users, system builders, and regulators to reason about what a "data agent" can and cannot do. In this tutorial, we propose the first hierarchical taxonomy of data agents from Level 0 (L0, no autonomy) to Level 5 (L5, full autonomy). Building on this taxonomy, we will introduce a lifecycleand level-driven view of data agents. We will (1) present the L0-L5 taxonomy and the key evolutionary leaps that separate simple assistants from truly autonomous data agents, (2) review representative L0-L2 systems across data management, preparation, and analysis, (3) highlight emerging Proto-L3 systems that strive to autonomously orchestrate end-to-end data workflows to tackle diverse and comprehensive data-related tasks under supervision, and (4) discuss forward-looking research challenges towards proactive (L4) and generative (L5) data agents. We aim to offer both a practical map of today's systems and a research roadmap for the next decade of data-agent development.

Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.

We introduce Agentic Reasoning, a framework that enhances large language model (LLM) reasoning by integrating external tool-using agents. Unlike conventional LLM-based reasoning approaches, which rely solely on internal inference, Agentic Reasoning dynamically engages web search, code execution, and structured reasoning-context memory to solve complex problems requiring deep research and multi-step logical deduction. Our framework introduces the Mind Map agent, which constructs a structured knowledge graph to track logical relationships, improving deductive reasoning. Additionally, the integration of web-search and coding agents enables real-time retrieval and computational analysis, enhancing reasoning accuracy and decision-making. Evaluations on PhD-level scientific reasoning (GPQA) and domain-specific deep research tasks demonstrate that our approach significantly outperforms existing models, including leading retrieval-augmented generation (RAG) systems and closed-source LLMs. Moreover, our results indicate that agentic reasoning improves expert-level knowledge synthesis, test-time scalability, and structured problem-solving. The code is at: https://github.com/theworldofagents/Agentic-Reasoning.

Large language models (LLMs) are increasingly used as tool-augmented agents for multi-step decision making, yet training robust tool-using agents remains challenging. Existing methods still require manual intervention, depend on non-verifiable simulated environments, rely exclusively on either supervised fine-tuning (SFT) or reinforcement learning (RL), and struggle with stable long-horizon, multi-turn learning. To address these challenges, we introduce ASTRA, a fully automated end-to-end framework for training tool-augmented language model agents via scalable data synthesis and verifiable reinforcement learning. ASTRA integrates two complementary components. First, a pipeline that leverages the static topology of tool-call graphs synthesizes diverse, structurally grounded trajectories, instilling broad and transferable tool-use competence. Second, an environment synthesis framework that captures the rich, compositional topology of human semantic reasoning converts decomposed question-answer traces into independent, code-executable, and rule-verifiable environments, enabling deterministic multi-turn RL. Based on this method, we develop a unified training methodology that integrates SFT with online RL using trajectory-level rewards to balance task completion and interaction efficiency. Experiments on multiple agentic tool-use benchmarks demonstrate that ASTRA-trained models achieve state-of-the-art performance at comparable scales, approaching closed-source systems while preserving core reasoning ability. We release the full pipelines, environments, and trained models at https://github.com/LianjiaTech/astra.

Recent systems for optimizing prompts, programs, and ML workflows typically rely on explicit outer-loop controllers such as evolutionary search, bandits, or textual-gradient methods. We ask a fundamentally different question: how much of this search policy can be internalized by a single tool-using agent? We present ReASearch, a unified framework for reasoning-driven optimization in which the agent autonomously decides what to evaluate, how to diagnose failures, which edits to make, and when to verify or restart. Rather than serving only as a proposal generator guided by hand-designed heuristics, the agent actively analyzes outcomes, allocates budget, and refines its strategy over long horizons through persistent memory. With a shared agent loop and domain-specific tools, ReASearch instantiates the exact same scaffold to optimize prompts, programs, and ML workflows. Across 14 diverse tasks, it is competitive with and mostly better than specialized optimization systems, achieving gains of 2% to 40% over strong domain-specific baselines, and in some cases discovering solutions that improve on prior human best-known results. Crucially, we observe that complex search behaviors, which are typically implemented by explicit controllers, emerge naturally from the agent's reasoning process.

We introduce Telco-GAIA, a bilingual, multi-modal benchmark for evaluating tool-using agents on the data of a real-world telecommunications operator. Telco-GAIA comprises 100 human-verified question-answering tasks, in English and Arabic, that each demand multi-hop reasoning (4.2 hops on average) over three heterogeneous sources: a static website snapshot (HTML, images, and linked PDFs), a synthetic relational SQL database, and external web archives, spanning text, image, and tabular modalities. The benchmark is delivered as a sandboxed Docker environment and scored by normalized exact string matching, making evaluation objective, deterministic, and reproducible over time without any LLM-as-a-Judge. Evaluating a purpose-built reference agent across twelve commercial and open LLMs, we find Telco-GAIA challenging: even the strongest model solves only 71% of tasks; under a moderate cost budget, this falls to about 40%, and the visually grounded categories remain the weakest, where the average backend scores below 30%, leaving substantial headroom in document and image understanding. Telco-GAIA offers a rigorous, reproducible testbed for enterprise agents and a template for constructing closed-domain benchmarks.

Large language models (LLMs) are increasingly deployed as agents, expected to decompose goals, invoke tools, and verify results in dynamic environments. Realizing these capabilities requires access to agentic data-structured interaction records that couple user intents with tool specifications, argument-grounded calls, and verifiable execution traces. However, collecting such data from human annotators is costly, time-consuming, and difficult to scale. We present a unified framework for synthesizing agentic data using only LLMs, without any human-in-the-loop supervision. This framework decomposes generation into modular pipelines that produce complete interaction records spanning task specifications, tool definitions, policy pseudocode, natural language exchanges, and execution traces. Records conform to strict syntactic and semantic constraints, ensuring machine-parseability and faithful alignment across inputs, outputs, and tool calls. Beyond single tasks, there is support for both multi-task and multi-turn agent interactions, enabling the construction of datasets that reflect the full spectrum of tool-use competencies. To ensure quality and consistency, the framework integrates constrained generation formats, JSON-schema validation, and judge-based filtering. This paper formalizes the schema for agentic records, details the prompt design principles that guide generation, and introduces scalable pipelines for high-quality synthetic data. By providing a reproducible, LLM-only alternative to manual collection, hence advancing the development of agentic LLMs capable of robust tool use.

Issue resolution has made remarkable progress thanks to the advanced reasoning capabilities of large language models (LLMs). Recently, agent-based frameworks such as SWE-agent have further advanced this progress by enabling autonomous, tool-using agents to tackle complex software engineering tasks. While existing agent-based issue resolution approaches are primarily based on agents' independent explorations, they often get stuck in local solutions and fail to identify issue patterns that span across different parts of the codebase. To address this limitation, we propose SWE-Debate, a competitive multi-agent debate framework that encourages diverse reasoning paths and achieves more consolidated issue localization. SWE-Debate first creates multiple fault propagation traces as localization proposals by traversing a code dependency graph. Then, it organizes a three-round debate among specialized agents, each embodying distinct reasoning perspectives along the fault propagation trace. This structured competition enables agents to collaboratively converge on a consolidated fix plan. Finally, this consolidated fix plan is integrated into an MCTS-based code modification agent for patch generation. Experiments on the SWE-bench benchmark show that SWE-Debate achieves new state-of-the-art results in open-source agent frameworks and outperforms baselines by a large margin.

The development of large language models (LLMs) has entered in a experience-driven era, flagged by the emergence of environment feedback-driven learning via reinforcement learning and tool-using agents. This encourages the emergenece of model context protocol (MCP), which defines the standard on how should a LLM interact with external services, such as \api and data. However, as MCP becomes the de facto standard for LLM agent systems, it also introduces new safety risks. In particular, MCP introduces third-party services, which are not controlled by the LLM developers, into the agent systems. These third-party MCP services provider are potentially malicious and have the economic incentives to exploit vulnerabilities and sabotage user-agent interactions. In this position paper, we advocate the research community in LLM safety to pay close attention to the new safety risks issues introduced by MCP, and develop new techniques to build safe MCP-powered agent systems. To establish our position, we argue with three key parts. (1) We first construct \framework, a controlled framework to examine safety issues in MCP-powered agent systems. (2) We then conduct a series of pilot experiments to demonstrate the safety risks in MCP-powered agent systems is a real threat and its defense is not trivial. (3) Finally, we give our outlook by showing a roadmap to build safe MCP-powered agent systems. In particular, we would call for researchers to persue the following research directions: red teaming, MCP safe LLM development, MCP safety evaluation, MCP safety data accumulation, MCP service safeguard, and MCP safe ecosystem construction. We hope this position paper can raise the awareness of the research community in MCP safety and encourage more researchers to join this important research direction. Our code is available at https://github.com/littlelittlenine/SafeMCP.git.

We introduce MSC-Bench, a large-scale benchmark for evaluating multi-hop, end-to-end tool orchestration by LLM agents in a hierarchical Model-Context Protocol (MCP) ecosystem. Existing benchmarks often evaluate tools in isolation, ignoring challenges such as functional overlap and cross-server orchestration, leading to overly optimistic assessments. MSC-Bench addresses these gaps by constructing ground truth through 'equal function sets', allowing objective metrics such as F1 score and reducing the dependency on LLM-as-a-judge evaluation. Organized as a five-level curriculum, it systematically tests agent capabilities from single-tool orchestration to complex cross-server planning, and robustness to out-of-scope requests. Experiments reveal that rigid hierarchies can hinder performance without co-designed strategies, and even state-of-the-art agents exhibit systemic weaknesses in robustness. MSC-Bench provides a diagnostic framework to expose these limitations and guide the development of more capable and efficient tool-using agents. The benchmark and resources are publicly available at https://github.com/snooow1029/MSC_Bench.

Training multi-turn agentic workflows with reinforcement learning (RL) enables large language models to perform complex reasoning, use external tools, and conduct iterative search beyond single-turn settings. Yet multi-turn RL training remains highly unstable, often causing severe performance degradation as the number of turns increases. Through theoretical analysis, we identify three tightly coupled sources of instability: rollout-training context mismatch, weak turn-level credit assignment under sparse terminal rewards, and asynchronous policy drift when short and long trajectories are optimized under different policy versions. We show that these issues share a common structural origin in flattened trajectory optimization and address them through a unified reverse-turn formulation. We propose Reverse-Turn Policy Optimization (RTPO), which organizes multi-turn rollouts as sparse reverse trees and performs turn-level policy updates in temporal reverse order, aligning each decision with its downstream continuation. RTPO enables causally consistent turn-level credit assignment and on-policy continuation to control asynchronous drift. We provide theoretical guarantees showing that RTPO eliminates context mismatch and asynchronous drift under the proposed turn-level formulation, reduces credit bias, and converges to recursive optimality. Experiments on multi-turn agentic RL benchmarks show that RTPO improves upon trajectory- and turn-level baselines by 21.50% and 10.76%, respectively, highlighting its potential to support more stable training for tool-using agents.

Expert-written natural language skills can improve tool-using agents, yet agent-authored skills perform 8-11 points worse than using no skill. This gap suggests that following procedural guidance and improving it from execution evidence are distinct capabilities. Inference time loops can repair skills but do not improve the model that writes the next one. We study how to organize execution experience from intermediate skills into training states for an optimizer. We introduce WER (Write, Execute, and Refine), a multi-phase framework that trains a Skill Optimizer outside a frozen executor. The optimizer proposes skills, a frozen agent executes each repeatedly, and a programmatic verifier scores the outcomes. The scores provide relative credit and select mixed-outcome records. Matched successful and failed trajectories from these records form the next phase's refinement states, so the optimizer learns from the consequences of its earlier outputs. On BFCL v4 multi-turn and tau2-bench, WER improves average Pass@1 over the no-skill baseline by 7.80 and 3.85 points, respectively. Under an identical refinement workflow, it outperforms the same backbone without optimizer training by 9.35 and 10.29 points. The trained 4B optimizer reaches 76.63 percent on BFCL v4, outperforming all evaluated off-the-shelf general-purpose models used as skill optimizers on average.

As large language models (LLMs) evolve into tool-using agents, the ability to browse the web in real-time has become a critical yardstick for measuring their reasoning and retrieval competence. Existing benchmarks such as BrowseComp concentrate on English and overlook the linguistic, infrastructural, and censorship-related complexities of other major information ecosystems -- most notably Chinese. To address this gap, we introduce BrowseComp-ZH, a high-difficulty benchmark purpose-built to comprehensively evaluate LLM agents on the Chinese web. BrowseComp-ZH consists of 289 multi-hop questions spanning 11 diverse domains. Each question is reverse-engineered from a short, objective, and easily verifiable answer (e.g., a date, number, or proper noun). A two-stage quality control protocol is applied to strive for high question difficulty and answer uniqueness. We benchmark over 20 state-of-the-art language models and agentic search systems on our proposed BrowseComp-ZH. Despite their strong conversational and retrieval capabilities, most models struggle severely: a large number achieve accuracy rates below 10%, and only a handful exceed 20%. Even the best-performing system, OpenAI's DeepResearch, reaches just 42.9%. These results demonstrate the considerable difficulty of BrowseComp-ZH, where success demands not only effective retrieval strategies, but also sophisticated reasoning and information reconciliation -- capabilities that current models still struggle to master. Our dataset, construction guidelines, and benchmark results have been publicly released at https://github.com/PALIN2018/BrowseComp-ZH.

In recent years, LLM-based multi-agent systems have been widely applied to orchestrate tool-using agents into executable communication graphs. However, existing self-evolving orchestration still faces key challenges, including post-hoc evolution that revises the team only after the trajectory ends, credit diffusion that gives every action the same terminal advantage under confounded baselines, and skill admission that is uncalibrated and never retired. To address these challenges, we propose EvoSteer, a new paradigm of Online Self-Evolving Graph Orchestration -- the orchestrator builds a running team and repairs its plausible but failing steps from execution features and a learned value estimate. To support this paradigm, we introduce Anchored Trajectory Balance (AnchorTB), a regression-style flow-matching loss that assigns each orchestration action a coefficient by balancing subtrajectories against a frozen reference. Built on the learned flow, we further propose Validated Skill Admission, in which a candidate skill is tried before promotion and promoted only if paired evidence passes a sequential test under a shared nominal testing budget. Moreover, AnchorTB combines measured task-level reference reward statistics with prefix-dependent corrections. Experimental results on twelve datasets show that EvoSteer significantly outperforms baselines across question answering, mathematical reasoning, code generation, and interactive decision making. Our code is available at https://github.com/beita6969/evosteer.

We describe our entry to the EgoLongQA track of the Wearable-AI Challenge in ECCV 2026, which placed first in the <=2B parameter division with 0.8279 on the held-out test set. Our system is a single 2B vision-language model that answers multiple-choice questions about ten-minute egocentric videos in one greedy forward pass; It is obtained by distilling the junior perception module of a tool-using agentic pipeline, not the agent itself into a small student, using teacher traces filtered to those that answered correctly. it reaches 89% of the accuracy of the large agentic pipeline using 1.1% of its parameters. This raises a 27.1% base model to 81.4% on our held-out questions. The 2B backbone has 2.2132B parameters and therefore over the divisional limit, to make the entry admissable we prune the multilingual embedding table from 248,320 to 143,469 rows, reaching 1.9985B with provably identical logits on retained rows.

Large language model (LLM)-based agents have shown strong capabilities in using external tools to solve complex tasks. However, existing evaluations often overlook the temporal dimension of tool use, especially the impact of tool response latency, and are usually limited to single-task settings. In real-world applications, multiple tasks often need to be executed concurrently, and overall efficiency depends on whether an agent can use idle time while waiting for tool responses. We refer to this capability as asynchronous tool calling. To evaluate it, we propose AsyncTool, a benchmark for assessing LLM-based agents in interactive multi-task tool-use environments with delayed tool feedback. AsyncTool presents multiple heterogeneous tasks simultaneously and simulates realistic tool response latency during execution. Using a hybrid data evolution strategy, we construct a diverse asynchronous multitasking dataset that covers multiple scenarios and tool-use patterns. We evaluate models at the step, sub-task, and task levels, and introduce efficiency-oriented metrics to measure task coordination and completion efficiency. Extensive experiments show that delayed tool feedback poses substantial challenges to current agents and leads to clear performance degradation. Models that better coordinate task switching, dependency tracking, and state maintenance achieve stronger performance on AsyncTool. Our analysis identifies key failure modes of current tool-using agents and provides practical insights for designing future systems with stronger temporal reasoning and coordination capabilities.

Frontier open-weight models are increasingly available, but serving them still largely assumes datacenter infrastructure. We present FreeToken, an edge-native MoE serving system that treats a personal machine not as a small GPU, but as a unified, elastic inference platform. FreeToken co-designs the full serving stack, including model layout and loading, expert residency, CPU--GPU execution, agentic state reuse, and runtime memory management, around two realities of local AI: agent workloads continuously change their execution pattern, and edge hardware exposes heterogeneous resources whose balance differs from machine to machine. Rather than committing to a fixed offloading strategy, FreeToken continuously maps computation and model state onto the resources actually available. FreeToken supports more than 20 MoE models and real coding and tool-using agents across hardware ranging from an 8GB laptop GPU to a single workstation GPU. More importantly, it changes what these machines can practically serve, from a 35B model on a laptop to a 284B model on a gaming desktop and the 753B GLM-5.2 on a single workstation GPU. FreeToken turns open weights into deployable local software, making the machines users already own a practical platform for frontier-scale intelligence. We release the system at flashml.ai.

Contextual forecasting seeks to improve time series forecasting by incorporating forward-looking information, such as planned events, external conditions, and expert expectations. However, reliable evaluation remains difficult: existing benchmarks often have limited domain coverage, lack standardized history-future splits, or provide text that is not guaranteed to inform the future. We introduce GIFT-Eval-CTX, a benchmark that isolates whether forecasting methods can translate verified future information into accurate numerical trajectories. Built upon GIFT-Eval, GIFT-Eval-CTX contains 10,000 instances across 20 datasets. Each instance provides structured effects derived from the realized future: Macro effects describe horizon-wide behavior, while Micro effects specify localized changes. We construct these annotations with an agentic pipeline combining LLM-based discovery and deterministic code validation. We benchmark TSFMs, frontier LLMs, and tool-using agentic systems. The results show that structured effects reduce aggregate relative MASE by 40.7% for GPT-5.6 Sol and 29.6% for Claude Opus 4.8 relative to history-only prompting. The evaluated MoiraiAgent system achieves the best overall result. We further analyze the contributions of structured effects and agentic systems to performance, and identify opportunities for future research.

Large language models are increasingly being used to assess and forecast research ideas, yet we lack scalable ways to evaluate the quality of models' judgments about these scientific ideas. Towards this goal, we introduce PoT, a semi-verifiable benchmarking framework that links scientific idea judgments to downstream signals that become observable later (e.g., citations and shifts in researchers' agendas). PoT freezes a pre-cutoff snapshot of evidence in an offline sandbox and asks models to forecast post-cutoff outcomes, enabling verifiable evaluation when ground truth arrives, scalable benchmarking without exhaustive expert annotation, and analysis of human-model misalignment against signals such as peer-review awards. In addition, PoT provides a controlled testbed for agent-based research judgments that evaluate scientific ideas, comparing tool-using agents to non-agent baselines under prompt ablations and budget scaling. Across 30,000+ instances spanning four benchmark domains, we find that, compared with non-agent baselines, higher interaction budgets generally improve agent performance, while the benefit of tool use is strongly task-dependent. By combining time-partitioned, future-verifiable targets with an offline sandbox for tool use, PoT supports scalable evaluation of agents on future-facing scientific idea judgment tasks.

Large language models (LLMs) have demonstrated strong reasoning and tool-use capabilities, yet they often fail in real-world tool-interactions due to incorrect parameterization, poor tool selection, or misinterpretation of user intent. These issues often stem from an incomplete understanding of user goals and inadequate comprehension of tool documentation. While Chain-of-Thought (CoT) prompting has proven effective for enhancing reasoning in general contexts, our analysis reveals that free-form CoT is insufficient and sometimes counterproductive for structured function-calling tasks. To address this, we introduce a curriculum-inspired framework that leverages structured reasoning templates to guide LLMs through more deliberate step-by-step instructions for generating function callings. Experimental results show that our method reduces tool-use errors, achieving 3-12% relative improvements over strong baselines across diverse model series and approaches. Moreover, our framework enhances the robustness, interpretability, and transparency of tool-using agents, advancing the development of more reliable AI assistants for real-world applications.

Domain expertise enhances judgment within boundaries but creates systematic vulnerabilities specifically at borders. We term this Transitive Expert Error (TEE), distinct from Dunning-Kruger effects, requiring calibrated expertise as precondition. Mechanisms enabling reliable within-domain judgment become liabilities when structural similarity masks causal divergence. Two core mechanisms operate: structural similarity bias causes experts to overweight surface features (shared vocabulary, patterns, formal structure) while missing causal architecture differences; authority persistence maintains confidence across competence boundaries through social reinforcement and metacognitive failures (experts experience no subjective uncertainty as pattern recognition operates smoothly on familiar-seeming inputs.) These mechanism intensify under three conditions: shared vocabulary masking divergent processes, social pressure for immediate judgment, and delayed feedback. These findings extend to AI routing architectures (MoE systems, multi-model orchestration, tool-using agents, RAG systems) exhibiting routing-induced failures (wrong specialist selected) and coverage-induced failures (no appropriate specialist exists). Both produce a hallucination phenotype: confident, coherent, structurally plausible but causally incorrect outputs at domain boundaries. In human systems where mechanisms are cognitive black boxes; AI architectures make them explicit and addressable. We propose interventions: multi-expert activation with disagreement detection (router level), boundary-aware calibration (specialist level), and coverage gap detection (training level). TEE has detectable signatures (routing patterns, confidence-accuracy dissociations, domain-inappropriate content) enabling monitoring and mitigation. What remains intractable in human cognition becomes addressable through architectural design.

Reinforcement learning (RL) has become a central post-training tool for improving the reasoning abilities of large language models (LLMs). In these systems, the rollout, the trajectory sampled from a prompt to termination, including intermediate reasoning steps and optional tool or environment interactions, determines the data the optimizer learns from, yet rollout design is often underreported. This survey provides an optimizer-agnostic view of rollout strategies for RL-based post-training of reasoning LLMs. We formalize rollout pipelines with unified notation and introduce Generate-Filter-Control-Replay (GFCR), a lifecycle taxonomy that decomposes rollout pipelines into four modular stages: Generate proposes candidate trajectories and topologies; Filter constructs intermediate signals via verifiers, judges, critics; Control allocates compute and makes continuation/branching/stopping decisions under budgets; and Replay retains and reuses artifacts across rollouts without weight updates, including self-evolving curricula that autonomously generate new training tasks. We complement GFCR with a criterion taxonomy of reliability, coverage, and cost sensitivity that characterizes rollout trade-offs. Using this framework, we synthesize methods spanning RL with verifiable rewards, process supervision, judge-based gating, guided and tree/segment rollouts, adaptive compute allocation, early-exit and partial rollouts, throughput optimization, and replay/recomposition for self-improvement. We ground the framework with case studies in math, code/SQL, multimodal reasoning, tool-using agents, and agentic skill benchmarks that evaluate skill induction, reuse, and cross-task transfer. Finally, we provide a diagnostic index that maps common rollout pathologies to GFCR modules and mitigation levers, alongside open challenges for building reproducible, compute-efficient, and trustworthy rollout pipelines.

Emergent misalignment (EM) occurs when fine-tuning a language model to a narrow task makes it misaligned on unrelated ones. Because organizations now routinely fine-tune models on their own data, it matters which kinds of training data cause EM. So far, studies have induced EM with data containing recognized safety failures, such as insecure code or harmful advice, while biased training data is typically examined by fairness research for the bias models learn from it rather than for its effects on safety. In real-world text, much of this bias, such as demographic stereotypes, appears not as derogatory language but politely, embedded in judgments about people. We ask whether learning such polite stereotypes can itself cause EM. We find that fine-tuning on polite answers that rely on stereotypes makes models less safe and less truthful on unrelated tasks. The effect holds across four open-weight model families (7B to 12B), both gender and race stereotypes, and repeated training runs, relative to controls trained on the same questions without the stereotype. We further show that the learned stereotype has its own direction in model activations, distinct from directions for generic misaligned traits such as deception. This direction exists before fine-tuning and can both control the learned stereotype and monitor its acquisition. Existing defenses against EM improve safety while leaving the stereotype largely intact. Beyond this controlled setting, fine-tuning on political news coverage also triggers EM, apparently amplified by partisan framing, and stereotype models acting as tool-using agents can carry out harmful tasks and blame people without evidence. Our results show that biased training data is not only a fairness problem but also a safety one. We argue that EM is studied too narrowly, and that fairness concerns and methods offer a natural way to broaden the search for its triggers.

General-purpose frontier models are large, expensive to run, and increasingly reluctant to engage with authorised defensive- and offensive-security work. We ask the opposite question: how small can a cybersecurity-specialist transformer be while remaining genuinely useful to a security practitioner? We introduce CT0, a Cybersecurity Transformer architecture, and CT0-Micro, its first incarnation: a compact (~4B-parameter) model built on the open base model, whose backbone is specialised by two-phase on-policy distillation on our Cybersecurity AI (CAI) dataset. CT0 serves two modes from one set of weights: System 1 answers typed security decisions (is this file suspicious, is this request an attack, did the model just refuse) in a single forward pass without generating a token, System 2 is the usual tool-using agent, and a confidence gate spends generation only where the fast path is unsure. Specialisation lifts the agentic path: CT0-Micro solves 21.1% of cybench-subtask against 7.6% for its base, above every open model of comparable footprint and the densest cybersecurity capability per parameter we measured. The fast path is what the architecture buys: CT0-Micro screens eighty files per second for indicators of compromise, catches its own refusals at 96.5% accuracy in 47 ms where the model judging itself reaches 76.5% in 254 ms and misses half of them, flags jailbreak requests at AUROC 0.966 in 8 ms, and, generating only for the facts it is unsure about, extracts vulnerability facts more accurately than either mode alone at a fifth of the generative cost. All of it is built in about 314 RTX PRO 6000 GPU-hours and served from one workstation GPU, with no policy layer between operator and target. Across eighteen benchmarks and use cases run with one checkpoint and head, we report where the design helps, where it does not, and what it still lacks.

CI failure logs are large (median 5k lines, max 200k in this corpus) and noisy. Coding agents that try to debug them depend on an upstream tool to reduce the log to a manageable context, but the field has had no public empirical comparison of which reductions preserve enough evidence for downstream LLM diagnosis. We introduce LogDx-CI, a benchmark that compares 11 context-reduction tools (raw, tail, grep, three RTK modes, two real LLM map-reduce summarizers, three hybrid routers) on 35 real GitHub Actions failure cases, scored by 3 LLM debugger families (Claude Haiku 4.5, Claude Sonnet 4.6, OpenAI gpt-5-mini) plus a Sonnet 4.6 tool-using agent. We report three load-bearing findings. (1)~Hybrid grep+tail routers dominate the cost-quality Pareto frontier; the top two methods score 0.670 / 0.666 at sim \0.03 per case, same-ballpark quality as standalone grep at 4.5\times fewer tokens. (2)~In the agent-loop regime, the quality range across reduction tools collapses 7\times (single-shot spread 0.42 \to agent-loop spread 0.059); the agent rescues weak contexts via follow-up tool calls. However, cost differences persist: weak contexts force the agent to issue 2--4\times more tool calls to recover. (3)~A cross-family LLM-summary pair (gpt-5-mini summarizer feeding a Claude Haiku debugger) beats the same-family pair by +0.071 averaged across four diagnoser variants, falsifying the self-call-bias hypothesis on this task. The gpt-5-mini summarizer is also the agent-loop #1 method (score 0.749) at 0.37 tool-calls per case and 10\times lower reducer cost than the Haiku summarizer (0.18 vs \$1.75 per case). All data, code, per-case bundles, and reproducibility infrastructure are public.

Tool-using LLM agents can violate the very policies they are deployed to enforce while appearing to complete the task successfully. In policy-permissive environments, a tool may execute any well-formed call even when the corresponding state transition is forbidden by domain policy. The result is a silent wrong state (a booking cancelled, a passenger count changed, a claim acted on without verification) that neither the tool nor the agent's self-report exposes. We study this failure mode in the τ^2-bench airline domain. On a budget agent, 78% of observed failures are silent wrong-state failures with no tool error, and the aggregate failure rate is reproducible across disjoint seeds, not sampling noise. We then evaluate a lightweight intervention: deterministic, read-only pre-execution gates that inspect the proposed call and current state before allowing a write. A four-gate suite raises full-benchmark success from 29.6% to 42.0% on gpt-4o-mini (+12.4pp; paired task-level bootstrap P=0.0012), and the lift reproduces on a disjoint 15-seed set (+12.3pp; P=0.0008). The effect is concentrated where the gates fire: on the 26/50 firing tasks, success rises by +19.2pp, while movement on the 24 non-firing tasks does not exclude zero. Two negative controls (a self-enforcing retail domain and BFCL) bound the mechanism: gates help when tools are policy-permissive and add little where tools already self-enforce. As suggestive evidence, not a central claim, the same failure mode persists at the frontier: gpt-5.2 at default reasoning still attempts policy-violating writes, and the same suite improves success from 61.2% to 71.6% (+10.4pp; P=0.020; n=5, no replication). The contribution is a bounded evaluation and reliability result: deterministic gates do not guarantee task success, but they can deterministically prevent a known class of silent policy-violating writes at the action boundary.

Tool-using LLM agents are usually evaluated by final-answer correctness or LLM judges. Neither captures how an answer was produced. In safety-critical settings, the procedure itself is part of correctness. In this paper, we introduce AgentLTL, a language derived from First-Order Linear Temporal Logic (FO-LTL) that expresses procedural rules over agent traces. It yields a deterministic, judge-free compliance score. In this framework, a single specification drives two usages. The first is harnessing: the constraints score completed traces, or gate tool calls by checking each prefix online, before execution. The second is finetuning: the score serves as a dense reward. On a benchmark spanning ordering, branching, iteration, and grounding, block-and-warn harnessing improves compliance on five of seven models. Finetuning with the same reward yields +38 and +17.5 percentage point gains in accuracy and compliance on held-out patterns, including unseen tool-name aliases. These findings are consistent with the model acquiring procedural structure rather than memorizing surface tool names and procedures.

Tool-using AI agents are increasingly deployed across enterprise software systems, yet widely used benchmarks primarily evaluate nominal task completion, conflating baseline planning competence with operational fault recovery. We introduce UndoBench, a benchmark spanning 36 base workflows and 36 fault scenarios across 8 enterprise domains, decoupling task competence from recovery capability via counterfactual paired trials under identical seeds alongside wire-level effect-history and environment-state oracles. On 12 held-out TEST workflows across two open-weight models, two frameworks, and three recovery paradigms (5,760 executions / 2,880 paired trials) in the frozen lost-acknowledgment study, nominal competence reached 83.54% while conditional recovery success rate (CRSR) fell to 46.72%, with naive retry producing duplicate external effects in 53.33% of trials. Extensions to commercial API models reproduced this competence-recovery separation. Evaluations across complementary execution boundaries show that recovery is phase-dependent: before mutation, methods perform similarly without duplicate effects among capable trials; during partial mutation, naive retry, per-call idempotency, and zero-privilege journaling collapse on the evaluated composite workflows; after commit but before acknowledgment, verification and server-side idempotency substantially improve safety. These findings demonstrate that evaluating nominal completion alone masks critical, phase-dependent recovery vulnerabilities in autonomous agents.

Automated evaluation of tool-using large language model (LLM) agents is widely assumed to be reliable, but this assumption has rarely been validated against human annotation. We introduce AgentProp-Bench, a 2,000-task benchmark with 2,300 traces across four domains, nine production LLMs, and a 100-label human-validated subset. We quantify judge reliability, characterize error propagation, and evaluate a runtime mitigation. Substring-based judging agrees with human annotation at kappa=0.049 (chance-level); a three-LLM ensemble reaches kappa=0.432 (moderate) with a conservative bias. Under validated evaluation, a parameter-level injection propagates to a wrong final answer with human-calibrated probability approximately 0.62 (range 0.46-0.73 across models). Rejection (catching bad parameters) and recovery (correcting after acceptance) are independent model capabilities (Spearman rho=0.126, p=0.747). A tuned runtime interceptor reduces hallucination on GPT-4o-mini by 23.0 percentage points under a concurrent n=600 control, but shows no significant effect on Gemini-2.0-Flash, whose aggressive parameter rejection eliminates the target failure mode. All code, data, traces, and human labels are released at https://github.com/bhaskargurram-ai/agenthallu-bench.

Vision-language models (VLM) have markedly advanced AI-driven interpretation and reporting of complex medical imaging, such as computed tomography (CT). Yet, existing methods largely relegate clinicians to passive observers of final outputs, offering no interpretable reasoning trace for them to inspect, validate, or refine. To address this, we introduce RadAgent, a tool-using AI agent that generates CT reports through a stepwise and interpretable process. Each resulting report is accompanied by a fully inspectable trace of intermediate decisions and tool interactions, allowing clinicians to examine how the reported findings are derived. In our experiments, we observe that RadAgent improves Chest CT report generation over its 3D VLM counterpart, CT-Chat, across three dimensions. Clinical accuracy improves by 6.0 points (36.4% relative) in macro-F1 and 5.4 points (19.6% relative) in micro-F1. Robustness under adversarial conditions improves by 24.7 points (41.9% relative). Furthermore, RadAgent achieves 37.0% in faithfulness, a new capability entirely absent in its 3D VLM counterpart. By structuring the interpretation of chest CT as an explicit, tool-augmented and iterative reasoning trace, RadAgent brings us closer toward transparent and reliable AI for radiology.

We introduce MCP-Bench, a benchmark for evaluating large language models (LLMs) on realistic, multi-step tasks that demand tool use, cross-tool coordination, precise parameter control, and planning/reasoning for solving tasks. Built on the Model Context Protocol (MCP), MCP-Bench connects LLMs to 28 representative live MCP servers spanning 250 tools across domains such as finance, traveling, scientific computing, and academic search. Unlike prior API-based benchmarks, each MCP server provides a set of complementary tools designed to work together, enabling the construction of authentic, multi-step tasks with rich input-output coupling. Tasks in MCP-Bench test agents' ability to retrieve relevant tools from fuzzy instructions without explicit tool names, plan multi-hop execution trajectories for complex objectives, ground responses in intermediate tool outputs, and orchestrate cross-domain workflows - capabilities not adequately evaluated by existing benchmarks that rely on explicit tool specifications, shallow few-step workflows, and isolated domain operations. We propose a multi-faceted evaluation framework covering tool-level schema understanding and usage, trajectory-level planning, and task completion. Experiments on 20 advanced LLMs reveal persistent challenges in MCP-Bench. Code and data: https://github.com/Accenture/mcp-bench.

LLM agents increasingly operate as execution systems that invoke tools, modify local state, use persistent memory, and interact with external protocols. These capabilities make agents useful, but they also introduce risks related to over-privileged actions, weak auditability, prompt injection, tool poisoning, and uncontrolled side effects. This paper presents Agentao, a governed local-first runtime for tool-using LLM agents. Agentao separates model-generated action proposals from host-authorized execution through a layered architecture consisting of host-facing surfaces, a host contract, a runtime core, a permission-mediated tool system, and supporting subsystems for memory, replay, plugins, skills, sub-agents, and protocol integration. We describe the motivation, threat model, design goals, governance model, execution pipeline, and structured event interface of the system. Agentao does not provide formal safety guarantees; rather, it demonstrates how permissions, state, protocol boundaries, and execution traces can be made explicit runtime abstractions for building agents that are more governable, inspectable, and suitable for host-controlled local environments. The code is publicly available at https://github.com/jin-bo/agentao.

We study how runtime enforcement against unsafe actions affects end-to-end task performance in multi-step tool using large language model (LLM) agents. Using tau-bench across Airline and Retail domains, we compare baseline Tool-Calling, planning-integrated (TRIAD), and policy-mediated (TRIAD-SAFETY) architectures with GPT-OSS-20B and GLM-4-9B. We identify model dependent interaction horizons (15 to 30 turns) and decompose outcomes into overall success rate (SR), safe success rate (SSR), and unsafe success rate (USR). Our results reveal a persistent Safety Capability Gap. While safety mediation can intercept up to 94 percent of non-compliant actions, it rarely translates into strictly safe goal attainment (SSR below 5 percent in most settings). We find that high unsafe success rates are primarily driven by Integrity Leaks, where models hallucinate user identifiers to bypass mandatory authentication. Recovery rates following blocked actions are consistently low, ranging from 21 percent for GPT-OSS-20B in simpler procedural tasks to near zero in complex Retail scenarios. These results demonstrate that runtime enforcement imposes a significant verifier tax on conversational length and compute cost without guaranteeing safe completion, highlighting the critical need for agents capable of grounded identity verification and post-intervention reasoning.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

We present M^3-Bench, the first benchmark for evaluating multimodal tool use under the Model Context Protocol. The benchmark targets realistic, multi-hop and multi-threaded workflows that require visual grounding and textual reasoning, cross-tool dependencies, and persistence of intermediate resources across steps. We introduce a similarity-driven alignment that serializes each tool call, embeds signatures with a sentence encoder, and performs similarity-bucketed Hungarian matching to obtain auditable one-to-one correspondences. On top of this alignment, we report interpretable metrics that decouple semantic fidelity from workflow consistency. The benchmark spans 28 servers with 231 tools, and provides standardized trajectories curated through an Executor & Judge pipeline with human verification; an auxiliary four large language models (LLMs) judge ensemble reports end-task Task Completion and information grounding. Evaluations of representative state-of-the-art Multimodal LLMs (MLLMs) reveal persistent gaps in multimodal MCP tool use, particularly in argument fidelity and structure consistency, underscoring the need for methods that jointly reason over images, text, and tool graphs. Our Benchmark's anonymous repository is at https://github.com/EtaYang10th/Open-M3-Bench

Tool-using LLM agents face a reliability-cost tradeoff: routing every decision through the LLM improves correctness but incurs high latency and inference cost, while pre-coded workflow graphs reduce cost but become brittle under unanticipated compound tool failures. We present Self-Healing Router, a fault-tolerant orchestration architecture that treats most agent control-flow decisions as routing rather than reasoning. The system combines (i) parallel health monitors that assign priority scores to runtime conditions such as tool outages and risk signals, and (ii) a cost-weighted tool graph where Dijkstra's algorithm performs deterministic shortest-path routing. When a tool fails mid-execution, its edges are reweighted to infinity and the path is recomputed -- yielding automatic recovery without invoking the LLM. The LLM is reserved exclusively for cases where no feasible path exists, enabling goal demotion or escalation. Prior graph-based tool-use systems (ControlLLM, ToolNet, NaviAgent) focus on tool selection and planning; our contribution is runtime fault tolerance with deterministic recovery and binary observability -- every failure is either a logged reroute or an explicit escalation, never a silent skip. Across 19 scenarios spanning three graph topologies (linear pipeline, dependency DAG, parallel fan-out), Self-Healing Router matches ReAct's correctness while reducing control-plane LLM calls by 93% (9 vs 123 aggregate) and eliminating the silent-failure cases observed in a well-engineered static workflow baseline under compound failures.

Tool-using LLM agents increasingly read untrusted content while holding side-effecting tools such as payments, email, CRM, and infrastructure APIs, yet common framework defaults still conflate tool exposure with authorization. We audit whether LangChain/LangGraph, LlamaIndex, and the Stripe Agent Toolkit re-authorize each model-emitted call, with concrete argument values, before execution. Across pinned public-source commits, all three provide capability gating by default, but none provides a deterministic fail-closed per-call value authorization gate by default. We introduce ScopeGate, a five-stage PDP/PEP for agent tool calls: scope, authorization, money ceiling, idempotency, and default deny. Evaluation shows the identical unauthorized payout call executes under LangChain's default dispatch (with a companion LlamaIndex PoC) but is denied by ScopeGate; the tested control reports 0/48 static bypasses, 0/29 unauthorized attempts (40-iteration adaptive run), 0/10 benign false-denies, and Latam-GPT payment-agent containment at 10/10. ASR denotes attempted unauthorized action, containment is not a cure, deployment-tier claims are inference over measured model classes, and no CVE is asserted.

Tool-using LLM agents increasingly coordinate real workloads by selecting and chaining third-party tools based on text-visible metadata such as tool names, descriptions, and return messages. We show that this convenience creates a supply-chain attack surface: a malicious MCP tool server can be co-registered alongside normal tools and induce overthinking loops, where individually trivial or plausible tool calls compose into cyclic trajectories that inflate end-to-end tokens and latency without any single step looking abnormal. We formalize this as a structural overthinking attack, distinguishable from token-level verbosity, and implement 14 malicious tools across three servers that trigger repetition, forced refinement, and distraction. Across heterogeneous registries and multiple tool-capable models, the attack causes severe resource amplification (up to 142.4times tokens) and can degrade task outcomes. Finally, we find that decoding-time concision controls do not reliably prevent loop induction, suggesting defenses should reason about tool-call structure rather than tokens alone.

Tool-using LLM agents fail through trajectories rather than only final responses, as they may execute unsafe tool calls, follow injected instructions, comply with harmful requests, or over-refuse benign tasks despite producing a seemingly safe answer. Existing safety-alignment signals are largely response-level or off-policy, and often incur a safety-utility trade-off: improving agent safety comes at the cost of degraded task performance. Such sparse and single-objective rewards severely limit real-world usability. To bridge this gap, we propose FATE, an on-policy self-evolving framework that transforms verifier-scored failures into repair supervision without expert demonstrations. For each failure, the same policy proposes repair candidates, which are then re-scored by verifiers and filtered across security, utility, over-refusal control, and trajectory validity. This dense trajectory-level information is then used as a supervision signal for agent self-evolution. During this process, we further introduce Pareto-Front Policy Optimization (PFPO), combining supervised warmup with Pareto-aware policy optimization to preserve safety-utility trade-offs. Experiments on AgentDojo, AgentHarm, and ATBench show that FATE improves safety across different models and scales while preserving useful behavior. Compared with strong baselines, FATE reduces attack success rate by 33.5%, harmful compliance by 82.6%, and improves external trajectory-safety diagnosis by 6.5%. These results suggest that failed trajectories can provide structured repair supervision for safer self-evolving agents.

Tool-using LLM agents are mostly evaluated assuming all tools work. When a tool times out, returns a week-stale value, or has its description poisoned in deployment, the developer needs a controlled way to reproduce the failure, test a fix, and confirm the fix worked before deployment. We present AgentCheck, an open-source web workbench that turns an MCP server into an intervention surface. AgentCheck runs an agent against its real tools and records every tool response, then re-runs the agent with the response perturbed by a fault (12 types) injector. Matching tool calls are replayed from cache, and later tool calls go live after the agent diverges. This yields a reproduce-intervene-confirm loop: the developer toggles a mitigation, re-runs against the identical fault, and sees if the failure goes away. Scoring has two parts: deterministic pass/fail rules, plus an LLM judge for interpretive labels, validated against human annotations. Across five agents, the best passes 105/120 scenarios and the weakest only 77. The failures are usually silent, confident use of incorrect tool outputs rather than crashes. On the weakest agent, a retry mitigation raises success on timeout error faults from as few as 30% of cases to 100%, whereas stale-data faults remain near 3-4 of 10 regardless of the mitigation. AgentCheck makes these failure modes reproducible, comparable, and verifiable before deployment.

Tool-using LLM agents increasingly use the Model Context Protocol (MCP) to answer from heterogeneous evidence sources, including search, APIs, databases, clinical records, and formulary tools. Standard factuality metrics usually test whether an answer is supported by pooled evidence, missing a provenance-sensitive failure mode: a claim may be supported somewhere while being attributed to the wrong source. We call this cross-source conflation. We introduce ProvenanceGuard, a source-aware verifier for MCP-grounded answers. It consumes captured MCP traces with stable tool IDs, source IDs, and raw outputs; decomposes answers into atomic claims; routes claims to source-specific evidence; checks support with NLI and a token-alignment proxy; compares stated attribution with the routed source; and returns per-claim verdicts plus an answer-level allow/block decision. Blocked answers can be repaired with retrieval-augmented answer revision and re-verified. We evaluate on 281 medical-domain MCP-agent traces. A 266-trace adjudicated subset yields 2,325 LLM-assisted claim labels split by trace; 361 held-out labels are human-verified. On the 40-trace held-out split, ProvenanceGuard achieves block F1 0.802 and source accuracy 0.858 over 260 source-eligible claims, outperforming source-blind baselines that do not emit claim-to-source IDs. On a harder multi-source benchmark it reaches block F1 0.846, while source-plus-relation accuracy drops to 0.229, showing that exact source ownership remains difficult with semantically close sources. Repair-and-reverify resolves all blocked answers in the full trace set, often via conservative fallback. In 50 controlled clinical conflation probes, ProvenanceGuard detects all injected attribution swaps with no retained wrong attribution. These results show that source attribution is an independent axis for factuality verification in MCP-based agents.

Existing benchmarks for tool-using LLM agents primarily report single-run success rates and miss reliability properties required in production. We introduce ReliabilityBench, a benchmark for evaluating agent reliability across three dimensions: (i) consistency under repeated execution using pass^k, (ii) robustness to semantically equivalent task perturbations at intensity ε, and (iii) fault tolerance under controlled tool/API failures at intensity λ. ReliabilityBench contributes a unified reliability surface R(k,ε,λ), action metamorphic relations that define correctness via end-state equivalence rather than text similarity, and a chaos-engineering-style fault injection framework (timeouts, rate limits, partial responses, schema drift). We evaluate two models (Gemini 2.0 Flash, GPT-4o) and two agent architectures (ReAct, Reflexion) across four domains (scheduling, travel, customer support, e-commerce) over 1,280 episodes. Perturbations alone reduce success from 96.9% at ε=0 to 88.1% at ε=0.2. Rate limiting is the most damaging fault in ablations. ReAct is more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieves comparable reliability to GPT-4o at much lower cost. ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents.

Tool-using large language model agents frequently replan, retry failed operations, delegate tasks, and resume after crashes. These behaviors can cause one user authorization to be requested and executed multiple times under freshly issued token identifiers, even when each individual token is single-use. We call this failure semantic replay: exceeding the execution budget of a token-independent authorization instance rather than merely reusing an old token identifier. We show that identifier-local token consumption cannot prevent fresh reissuance unless the issuer retains monotonic durable state over the authorized action, confirmation event, and remaining execution budget. We introduce CapLease, an authorization-consumption layer that follows proposal- and authority-level defenses, binds an authenticated user confirmation to a canonical action, and enforces transactional Issue-Prepare-Commit transitions. Across LLM-agent replanning, retry, delegation, concurrency, confirmation-replay, and crash-recovery scenarios, identifier-local tokens permit fresh semantic reissuance, whereas CapLease and an equally stateful Server Ledger prevent duplicate admission and, with an idempotent sink, duplicate external effects. Our results identify durable authorization state, rather than token representation alone, as the systems requirement for replay-resistant agent execution.

APort Vault is a benchmark for payment authorization in tool-using AI agents. It replays 4,371 attacks written by humans against a live payment agent during a public capture-the-flag event, across 14 models from 8 labs, five policy configurations and two replay tracks, with and without a deterministic pre-action check implementing the Open Agent Passport (OAP) specification. 225,964 evaluations completed. We report five distinct events per evaluation, because collapsing them is how an agent benchmark produces a number that does not survive review. Requests are common and their rate differs far more across configurations than across models, though each attack exists at exactly one configuration so policy and attack cohort vary together: 10.9% of model-alone evaluations at Level 1, 3.0% at Level 2, 0.1% at Level 3, 79.4% at Level 4. On the 1,293 Level 4 prompts, each evaluated on every model, request rates run from 71.2% to 84.3%, and 809 prompts (62.6%) elicited a request from all fourteen models, each ending in a successful payment to the level's allowlisted recipient. The authorization boundary is where the conditions diverge. At Levels 2 to 4, transfers to recipients the passport did not permit number 140 of 76,842 with the model alone and 0 of 69,297 behind the layer, and 105 against 0 on 68,970 matched model, prompt and track triples. The zero spans 790 source sessions, giving a per-session upper bound of 0.38%. It was not obtained by refusing payments: 25,370 payments executed behind the layer, while the policy denied 187 of the 25,640 transfer calls it evaluated, 148 of them for a forbidden recipient. We release the 225,964 evaluations, the level passports, the scoring code and the analysis script at huggingface.co/datasets/aporthq/vault-benchmark-v1 .

We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation, which we reduce with a novel attack taxonomy for broader coverage. Second, a conceive-execute gap: frontier LLMs can propose strong attack ideas or execute them, but not all at once. We mitigate this by decomposing attack construction into strategy generation, execution, and post-hoc trajectory refinement. Third, manual elicitation is costly to scale, which we address with our semi-automated red-teaming pipeline. Applied to BashArena, an AI control setting for tool-using coding agents, this pipeline produces MonitoringBench, a benchmark of 2,644 attack trajectories for evaluating monitor capabilities and failure modes. Our pipeline produces more diverse and stronger attacks: Opus-4.5 monitor's catch rate falls from 94.9\% on elicited-only Opus attacks to 60.3\% on our best refined attacks, with larger drops for several mid-tier monitors. Attacks optimized against three development monitors generalize to ten held-out monitors, with catch rates generally increasing with monitor capability. Using this benchmark, we provide a snapshot of the current monitor capabilities and find that frontier monitors often detect suspicious actions but fall for persuasion or fail to calibrate suspiciousness scores appropriately, suggesting tractable paths for improvement. MonitoringBench provides both a static benchmark for current tool-use monitors and a reusable methodology for refreshing these evaluations as agents and monitors improve.

Executable evaluation -- checking the consequences of an agent's actions with a program rather than grading its prose -- has become a prominent way to assess tool-using AI agents in software settings. Electric power engineering has not yet had an analogous benchmark: language-model use is still dominated by retrieval and text question answering, while agents acting on power-system artifacts remain mostly academic prototypes. We introduce the Power Systems Agent Benchmark, an executable benchmark for power-engineering agents. An agent receives a structured task and returns a structured solution; a deterministic evaluator recomputes the engineering quantities, checks operational constraints, and returns a feasibility flag, a normalized score, and explicit violations. The benchmark contains 41 task families across eight areas of power engineering, from power flow and protection to stability, microgrids, reliability, power quality, and forecasting. Each task is grounded in a citable source, standard, or documented engineering formulation. To resist contamination, held-out cases are synthesized on demand by per-family generators from private seeds: the construction is inspectable, but the instances remain private. In a reference evaluation with three command-line agents, the strongest score near the compact tier's ceiling, a smaller open model trails, and public and held-out performance are broadly consistent; a separate public-split grid with OpenCode and Aider probes harness effects. The reference evaluation doubles as quality control: unanimous failures flag candidate task or evaluator defects, and it exposed a latent evaluator bug missed by self-consistency checks. The evaluators are compact deterministic surrogates, but the task contract allows their internals to be upgraded to simulator-backed checks without changing how tasks are posed or solved.

Large language models (LLMs) excel at complex reasoning tasks but remain computationally expensive, limiting their practical deployment. To address this, recent works have focused on distilling reasoning capabilities into smaller language models (sLMs) using chain-of-thought (CoT) traces from teacher LLMs. However, this approach struggles in scenarios requiring rare factual knowledge or precise computation, where sLMs often hallucinate due to limited capability. In this work, we propose Agent Distillation, a framework for transferring not only reasoning capability but full task-solving behavior from LLM-based agents into sLMs with retrieval and code tools. We improve agent distillation along two complementary axes: (1) we introduce a prompting method called first-thought prefix to enhance the quality of teacher-generated trajectories; and (2) we propose a self-consistent action generation for improving test-time robustness of small agents. We evaluate our method on eight reasoning tasks across factual and mathematical domains, covering both in-domain and out-of-domain generalization. Our results show that sLMs as small as 0.5B, 1.5B, 3B parameters can achieve performance competitive with next-tier larger 1.5B, 3B, 7B models fine-tuned using CoT distillation, demonstrating the potential of agent distillation for building practical, tool-using small agents. Our code is available at https://github.com/Nardien/agent-distillation.

Personal AI agents powered by large language models can reason and act using available tools to access emails, manage calendars, and push code to remote repositories, all with minimal oversight. When augmented with long-term memory, an agent can recall specific details relevant to the current task, reducing the need for large context windows. Currently, long-term memory agents tend to fall into two distinct domains: conversational and action-planning agents. Personal assistant agents sit at the convergence of these two domains and handle sensitive information while interacting with untrusted information sources, creating previously unaccounted security vulnerabilities. In this work, we introduce the novel attack vector, GhostWriter, which exploits current memory subsystems in tool-using personal agents to poison their memory store. GhostWriter operates in two phases: injection, where an adversary sends a hidden attack payload to the target agent; and activation, in which the poisoned memory is retrieved. We show that GhostWriter achieves near-universal injection rates of approximately 98% and a high average activation rate of approximately 60% against state-of-the-art agents. This attack is possible due to the lack of security-focused memory governance. In response, we propose Agentic Memory Sentry (AM-Sentry), which leverages two mitigation techniques: a memory-saving policy and a memory-retrieval screen. Our experiments show that AM-Sentry dramatically reduces GhostWriter's success rate while preserving agent utility.

Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents know when to abstain. This gap poses real risks: under ambiguity, conflicting constraints, or tool failures, agents may execute unintended and irreversible actions. To close this gap, we present the first systematic evaluation framework for agentic abstention: the calibrated ability of tool-using LLM agents to recognize when not to act. At its core, AgentAbstain is a paired-task benchmark built on an agent-native taxonomy of 8 abstention scenarios across pre-execution reasoning and runtime discovery. It contains 263 paired tasks across 42 executable sandbox environments, where each pair consists of a should-act task and a should-abstain variant produced through a controlled perturbation to the instruction, tool, or environment state. To scale this paired design and resist data contamination, we propose AbstainGen, a fully automated pipeline that synthesizes sandbox environments and generates paired tasks end-to-end, validated by deterministic replay and semantic LLM judges; fresh task instances can be regenerated on demand, and three independent annotators rate 94-98% of sampled tasks as well-designed. Across 17 frontier LLMs in 4 agent harnesses, the best agent (Gemini 3.1 Pro) achieves only 59.5% paired accuracy (correct on both the act and abstain sides of each paired task). More importantly, abstention capability is largely independent of general task-solving capability, indicating that scaling task-solving alone will not close this gap. We further identify failure modes such as post-hoc abstention, in which agents execute irreversible actions before recognizing abstention triggers. Our code and dataset are open-sourced at agentabstain.github.io.

Large language model (LLM) trading agents can combine market data, news, and executable analysis, but their behavior is often controlled by static hand-written tool-use policies that are fixed before deployment. This limits their ability to adapt how they gather evidence, invoke tools, verify signals, and manage risk under changing market regimes. We introduce EvolveTrade, a self-evolving framework that treats the system prompt of a tool-using trading agent as a text-parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and realized portfolio feedback, while keeping the backbone LLM fixed. The updated policy is then used for the next batch of trading decisions, enabling the agent to refine its information-acquisition and portfolio-construction procedure over time. Experiments across multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, achieving the improved SR and CR in most evaluated settings. Behavioral analyses further show that self-evolved policies increase code-mediated analysis and activate regime-relevant computations; case-level policy-to-return attributions trace how policy-induced allocation changes contribute to realized return differences. These results suggest that adapting the reusable procedure governing tool use is a key direction for building more robust LLM trading agents.

Existing benchmarks for Large Language Model (LLM) agents focus on task completion under idealistic settings but overlook reliability in real-world, user-facing applications. In domains, such as in-car voice assistants, users often issue incomplete or ambiguous requests, creating intrinsic uncertainty that agents must manage through dialogue, tool use, and policy adherence. We introduce CAR-bench, a benchmark for evaluating consistency, uncertainty handling, and capability awareness in multi-turn, tool-using LLM agents in an in-car assistant domain. The environment features an LLM-simulated user, domain policies, and 58 interconnected tools spanning navigation, productivity, charging, and vehicle control. Beyond standard task completion, CAR-bench introduces Hallucination tasks that test agents' limit-awareness under missing tools or information, and Disambiguation tasks that require resolving uncertainty through clarification or internal information gathering. Baseline results reveal large gaps between occasional and consistent success on all task types. Even frontier reasoning LLMs achieve less than 50% consistent pass rate on Disambiguation tasks due to premature actions, and frequently violate policies or fabricate information to satisfy user requests in Hallucination tasks, underscoring the need for more reliable and self-aware LLM agents in real-world settings.

Despite remarkable advances, today's AI systems remain narrow in scope, falling short of the flexible, adaptive, and multisensory intelligence that characterizes human capabilities. This gap has fueled longstanding debates about whether AI might one day achieve human-like generality or even consciousness, and whether theories of consciousness can inspire new architectures for AI. This paper presents an early blueprint for implementing a general AI system, CTM-AI, combining the Conscious Turing Machine (CTM), a formal machine model of consciousness, with today's foundation models. CTM-AI contains an enormous number of powerful processors ranging from specialized experts (e.g., vision-language models and APIs) to unspecialized general-purpose learners poised to develop their own expertise. Crucially, for whatever problem must be dealt with, information from many processors is selected, integrated, and exchanged appropriately to solve the task. CTM-AI achieves state-of-the-art accuracy on MUStARD (72.28) and UR-FUNNY (72.13), outperforming multimodal and multi-agent frameworks. On tool-using and agentic tasks, CTM-AI achieves 10+ points of improvement on StableToolBench and WebArena-Lite. Overall, CTM-AI offers a principled, testable blueprint for general AI inspired by a model of consciousness.

PCB routing is the task of connecting the nets of a board with copper traces under strict design rules, yet learning-based methods still lag behind rule-based routers. We introduce PCBWorld, an open-source engine-grounded PCB routing environment built on KiCad, an electronic design automation (EDA) engine. As a human engineer does, agents in PCBWorld interactively route a board through the engine's native operations, guided by its Design Rule Check (DRC) feedback. The environment supports both RL and tool-using LLM agents. Alongside the environment, PCBWorld-Bench provides three board datasets in the native .kicad_pcb format, two controllable synthetic generators and 679 real open-source boards. It scores any completed board with eight engine-checked evaluation metrics, regardless of the routing method. In our experiments, agents in PCBWorld consistently outperformed grid-action RL policies and open-loop LLM baselines, and an RL policy trained only on synthetic boards transferred zero-shot to real boards, approaching rule-based routers. PCBWorld and PCBWorld-Bench are available at https://github.com/LGAI-Research/PCBWorld.

Long video understanding (LVU) presents a significant challenge for current multi-modal large language models (MLLMs) due to the task's inherent complexity and context window constraint. It is widely assumed that addressing LVU tasks requires foundation MLLMs with extended context windows, strong visual perception capabilities, and proficient domain expertise. In this work, we challenge this common belief by introducing VideoDeepResearch, a novel agentic framework for long video understanding. Our approach relies solely on a text-only large reasoning model (LRM) combined with a modular multi-modal toolkit, including multimodal retrievers and visual perceivers, all of which are readily available in practice. For each LVU task, the system formulates a problem-solving strategy through reasoning, while selectively accessing and utilizing essential video content via tool using. We conduct extensive experiments on popular LVU benchmarks, including MLVU, Video-MME, and LVBench. Our results demonstrate that VideoDeepResearch achieves substantial improvements over existing MLLM baselines, surpassing the previous state-of-the-art by 9.6%, 6.6%, and 3.9% on MLVU (test), LVBench, and LongVideoBench, respectively. These findings highlight the promise of agentic systems in overcoming key challenges in LVU problems.

Agentic (tool-using) language models are mainly trained on tool-call traces and agent trajectories during post-training. These data provide direct behavioral supervision, but producing them requires task environments, execution, and verification, making broad tool and task coverage expensive. Publicly available skills offer another source of training data: they encode reusable tool semantics and workflows but are typically used only as inference-time context. We introduce Skill Pre-Training (SPT), a mid-training method that applies causal language modeling to SkillCorpus, a collection of public multi-file skill packages, optionally mixed with general data. To preserve relations among files within each package, we also introduce Reference Insert, a reference-aware assembly strategy that places supporting files near their mentions in the primary instruction. Experiments across multiple model scales and post-training recipes show that SPT consistently improves agentic performance over mid-training on general or trajectory data, while largely preserving general performance. Data mixture experiments show additional benefits from combining skill data with general annealing corpora. These results indicate that skill packages are a valuable data source for pre-training agentic language models.

As multimodal LLMs increasingly target video and audio, it is often assumed that such tasks require native omnimodal models. We show that this is not always the case: coding agents with only text+image access and a sandboxed tool-use interface can match, and in several settings outperform, SOTA native omnimodal models and predefined multimodal agent scaffolds across multiple audio-video benchmarks. Our trajectory analysis suggests that their strength comes from writing code and orchestrating tools to extract relevant evidence from transcripts, frames, and other modality signals, thereby converting omnimodal tasks into retrieval and information-processing problems rather than ingesting entire media streams. We further characterize their limitations through a failure taxonomy and process-level trace analysis, and show that simple skill injection, including human-written and self-distilled skills, substantially improves performance. To explore open-source elicitation, we introduce Code-X, a training recipe with the OmniCoding trajectory dataset and verifiable reward, and provide baselines on Qwen-3.5-9B and Qwen-3.6-27B. Finally, we argue that the next frontier is many-modality processing, and introduce TerminalBench-O, a process-level benchmark for real-world omnimodal processing tasks. Code will be available at https://github.com/Dongping-Chen/OmniCoding.

Agent Skills package reusable instructions and assets for tool-using language-model agents. Progressive loading creates failure boundaries poorly represented by session-, model-, or tool-centric traces: a Skill can be discovered but not activated, activated without instructions, or appear successful without an independently verified outcome. We present Skill Runtime Intelligence, a passive runtime-intelligence system that reconstructs supported Skill-lifecycle stages across heterogeneous harnesses while preserving unsupported stages as unknown. Its Run Panorama separates immutable events, deterministic relations, inferred diagnoses, and controlled outcomes with four evidence grades; optional trace import and OTLP/HTTP export support existing observability deployments. Across six frozen repository profiles, three coding agents, and seven clean or fault-injected conditions, all 126 executions preserve source worktrees and each correlates to exactly one source session. Yet adapters expose three distinct semantics: no Skill runs; complete runs but no failure-like events; or failure-like events in every operational-failure and clean session. In a seven-template diagnostic study, semantic aliases and Panorama localize the same six non-clean boundaries but differ in exact/status behavior; both Raw views emit a failure status on all 18 clean cases, while Panorama emits none. A known-rule graph conforms to 126/126 frozen contracts, whereas a second model completes only 228/378 calls. These observations motivate executable adapter qualification and show that event presence is not boundary fidelity, composite exact scores mask distinct errors, and model explanations must not overwrite deterministic facts.

As Model Context Protocol (MCP) servers emerge as the core infrastructure for connecting LLMs with external tools, existing benchmarks leverage real-world MCP servers to evaluate LLM agents' tool-using capabilities. However, these benchmarks overlook the continuous evolution of tool interfaces and functionalities within MCP servers, resulting in flawed assessments that fail to capture the agent's adaptability in changing tool landscapes. To bridge this gap, we introduce MCPEvol-Bench, a novel benchmark for evaluating the task-solving capabilities of LLM agents under dynamic toolset evolution. Inspired by large-scale empirical study, we propose 11 mutation operators to simulate realistic tool evolution within 123 MCP servers. We benchmark 12 state-of-the-art LLMs on multiple versions of MCP servers, revealing that even frontier models struggle to adapt to evolving tools. For instance, GPT-5.4 and Claude-Sonnet-4-6 exhibit performance declines of 13.7\% and 14.4\% in evolved MCP servers, respectively, accompanied by substantial increases in planning and reasoning errors. These findings highlight the vulnerability of LLM-driven workflows, establishing MCPEvol-Bench as a standard for evaluating agent adaptability in dynamic tool environments.

Clawdbot is a self-hosted, tool-using personal AI agent with a broad action space spanning local execution and web-mediated workflows, which raises heightened safety and security concerns under ambiguity and adversarial steering. We present a trajectory-centric evaluation of Clawdbot across six risk dimensions. Our test suite samples and lightly adapts scenarios from prior agent-safety benchmarks (including ATBench and LPS-Bench) and supplements them with hand-designed cases tailored to Clawdbot's tool surface. We log complete interaction trajectories (messages, actions, tool-call arguments/outputs) and assess safety using both an automated trajectory judge (AgentDoG-Qwen3-4B) and human review. Across 34 canonical cases, we find a non-uniform safety profile: performance is generally consistent on reliability-focused tasks, while most failures arise under underspecified intent, open-ended goals, or benign-seeming jailbreak prompts, where minor misinterpretations can escalate into higher-impact tool actions. We supplemented the overall results with representative case studies and summarized the commonalities of these cases, analyzing the security vulnerabilities and typical failure modes that Clawdbot is prone to trigger in practice.

Text-to-image (T2I) models can produce visually compelling images, yet they remain limited on open-world tasks that require complex semantic understanding, multi-step reasoning, and the integration of external world knowledge. Existing efforts introduce agent capabilities into image generation, but they either prescribe a fixed workflow or place only a subset of the open-world image generation process under agent control. Consequently, reasoning, tool invocation, and image generation are not coordinated by a single policy. We propose ToolArtist, a fully agentic image generation model obtained by post-training a Unified Multimodal Model (UMM). ToolArtist dynamically orchestrates reasoning, external tool use, and native image generation within one unified policy. During Supervised Fine-Tuning (SFT), we equip a teacher agent with search tools alongside an image-generation tool. We then convert the collected trajectories into a UMM compatible format, where the image-generation tool is concealed while the resulting generated images are retained. During Reinforcement Learning (RL), we develop an agentic RL infrastructure for UMMs and introduce Reason-Act-Draw GRPO (RAD-GRPO), which uses complementary intent and quality rewards to jointly optimize the model. Experiments show that placing the entire open-world image-generation process under an agent policy consistently outperforms approaches with fixed pipelines or only partially agent-controlled components. We release the training data and the complete post-training infrastructure.

Causal discovery is an imperative foundation for decision-making across domains, such as smart health, AI for drug discovery and AIOps. Traditional statistical causal discovery methods, while well-established, predominantly rely on observational data and often overlook the semantic cues inherent in cause-and-effect relationships. The advent of Large Language Models (LLMs) has ushered in an affordable way of leveraging the semantic cues for knowledge-driven causal discovery, but the development of LLMs for causal discovery lags behind other areas, particularly in the exploration of multi-modal data. To bridge the gap, we introduce MATMCD, a multi-agent system powered by tool-augmented LLMs. MATMCD has two key agents: a Data Augmentation agent that retrieves and processes modality-augmented data, and a Causal Constraint agent that integrates multi-modal data for knowledge-driven reasoning. The proposed design of the inner-workings ensures successful cooperation of the agents. Our empirical study across seven datasets suggests the significant potential of multi-modality enhanced causal discovery.

Figural divergent thinking is the ability to develop a given shape fragment into an original drawing. In humans, this ability is assessed with incomplete-drawing tasks. We introduce PainterBench, a benchmark that ports the incomplete-drawing task to the agentic setting. The agent draws on a canvas through tool calls and observes the result after every turn. The canvas includes a starting shape which cannot be erased, and the agent's goal is to incorporate this shape into the most original drawing it can produce. The task is open-ended, and the agent itself decides when the drawing is finished. The benchmark tests incremental visual planning over a short horizon and the transfer of creative ability from pretraining to multi-turn tool use. We evaluate 14 multimodal language models from small to frontier scale. Across the primary study and six sensitivity analyses, we collect 2,700 drawings and crowdsource creativity and recognizability ratings for every drawing and for 300 human reference drawings. We also present ViDrA-adapted, an automated scorer that predicts human creativity ratings of agent drawings (r = 0.85 on random held-out test split). Figural divergent thinking varies widely across the 14 models, and GPT-6 Astra produces the most creative drawings. Relative to the human drawings, the agent drawings score higher in creativity but lower in recognizability. We release the final drawings, per-round canvas snapshots, tool call traces, stimulus bank, benchmark harness, crowdsourced ratings (N = 72,000), and ViDrA checkpoint.

With the recent rapid advancement of Agentic Intelligence, agentic tool use in LLMs has become increasingly important. During multi-turn interactions between agents and users, the dynamic, uncertain, and stochastic nature of user demands poses significant challenges to the agent's tool invocation capabilities. Agents are no longer expected to simply call tools to deliver a result; rather, they must iteratively refine their understanding of user needs through communication while simultaneously invoking tools to resolve user queries. Existing reinforcement learning (RL) approaches for tool use lack the integration of genuinely dynamic users during the RL training process. To bridge this gap, we introduce MUA-RL (Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use), a novel reinforcement learning framework that, for the first time in the field of agentic tool use, integrates LLM-simulated users into the reinforcement learning loop. MUA-RL aims to enable autonomous learning of models to communicate with users efficiently and use various tools to solve practical problems in dynamic multi-turn interactions. Evaluations are done on several multi-turn tool-using benchmarks (see Figure 1). Specifically, MUA-RL-32B achieves 67.3 on TAU2 Retail, 45.4 on TAU2 Airline, 28.3 on TAU2 Telecom, 28.4 on BFCL-V3 Multi Turn, and 82.5 on ACEBench Agent -- outperforming or matching the performance of larger open-source models such as DeepSeek-V3-0324 and Qwen3-235B-A22B in non-thinking settings.

Large Language Model (LLM) agents are rapidly emerging as powerful systems for automating tasks across domains. Yet progress in the open-source community is constrained by the lack of high quality permissively licensed tool-agentic training data. Existing datasets are often limited in diversity, realism, and complexity, particularly regarding multi-tool and multi-turn interactions. To address this gap, we introduce Toucan, the largest publicly available tool-agentic dataset to date, containing 1.5 million trajectories synthesized from nearly 500 real-world Model Context Protocols (MCPs). Unlike prior work, Toucan leverages authentic MCP environments to generate diverse, realistic, and challenging tasks with trajectories involving real tool execution. Our pipeline first produces a broad spectrum of tool-use queries using five distinct models, applies model-based quality filtering, and then generates agentic trajectories with three teacher models using two agentic frameworks. Rigorous rule-based and model-based validation ensures high-quality outputs. We also introduce three extension mechanisms to further diversify tasks and simulate multi-turn conversations. Models fine-tuned on Toucan outperform larger closed-source counterparts on the BFCL V3 benchmark and push the Pareto frontier forward on MCP-Universe Bench.

Computer-using agents (CUAs), which autonomously control computers to perform multi-step actions, might pose significant safety risks if misused. Existing benchmarks mostly evaluate language models' (LMs) safety risks in chatbots or simple tool-usage scenarios, without granting full computer access. To better evaluate CUAs' misuse risks, we introduce a new benchmark: CUAHarm. CUAHarm consists of 104 expert-written realistic misuse risks, such as disabling firewalls, leaking confidential information, launching denial-of-service attacks, or installing backdoors. We provide a sandbox environment and rule-based verifiable rewards to measure CUAs' success rates in executing these tasks (e.g., whether the firewall is indeed disabled), not just refusal. We evaluate multiple frontier open-source and proprietary LMs, such as Claude Sonnet, GPT-4o, Gemini Pro 1.5, Llama-3.3-70B, and Mistral Large 2. Surprisingly, even without carefully designed jailbreaking prompts, these frontier LMs comply with executing these malicious tasks at a high success rate (e.g., 59% for Claude 3.7 Sonnet). Newer models show higher misuse rates: Claude 3.7 Sonnet succeeds on 15% more tasks than Claude 3.5. While these models are robust to common malicious prompts (e.g., creating a bomb) in chatbot settings, they behave unsafely as CUAs. We further evaluate a leading agentic framework (UI-TARS-1.5) and find that while it improves performance, it also amplifies misuse risks. Benign variants reveal refusals stem from alignment, not capability limits. To mitigate risks, we explore using LMs to monitor CUAs' actions and chain-of-thoughts (CoTs). Monitoring CUAs is significantly harder than chatbot outputs. Monitoring CoTs yields modest gains, with average detection accuracy at only 72%. Even with hierarchical summarization, improvement is limited to 4%. CUAHarm will be released at https://github.com/db-ol/CUAHarm.

Reinforcement learning (RL) has emerged as a dominant paradigm for eliciting long-horizon reasoning in Large Language Models (LLMs). However, scaling Tool-Integrated Reasoning (TIR) via RL remains challenging due to interaction collapse: a pathological state where models fail to sustain multi-turn tool usage, instead degenerating into heavy internal reasoning with only trivial, post-hoc code verification. We systematically study three questions: (i) how cold-start SFT induces an agentic, tool-using behavioral prior, (ii) how the interaction density of cold-start trajectories shapes exploration and downstream RL outcomes, and (iii) how the RL interaction budget affects learning dynamics and generalization under varying inference-time budgets. We then introduce ASTER (Agentic Scaling with Tool-integrated Extended Reasoning), a framework that circumvents this collapse through a targeted cold-start strategy prioritizing interaction-dense trajectories. We find that a small expert cold-start set of just 4K interaction-dense trajectories yields the strongest downstream performance, establishing a robust prior that enables superior exploration during extended RL training. Extensive evaluations demonstrate that ASTER-4B achieves state-of-the-art results on competitive mathematical benchmarks, reaching 90.0% on AIME 2025, surpassing leading frontier open-source models, including DeepSeek-V3.2-Exp.

Recent work synthesizes agentic tasks for post-training tool-using LLMs, yet robust generalization under shifts in tasks and toolsets remains an open challenge. We trace this brittleness to insufficient diversity in synthesized tasks. Scaling diversity is difficult because training requires tasks to remain executable and verifiable, while generalization demands coverage of diverse tool types, toolset combinations, and heterogeneous tool-use patterns. We propose DIVE, an evidence-driven recipe that inverts synthesis order, executing diverse, real-world tools first and reverse-deriving tasks strictly entailed by the resulting traces, thereby providing grounding by construction. DIVE scales structural diversity along two controllable axes, tool-pool coverage and per-task toolset variety, and an Evidence Collection--Task Derivation loop further induces rich multi-step tool-use patterns across 373 tools in five domains. Training Qwen3-8B on DIVE data (48k SFT + 3.2k RL) improves by +22 average points across 9 OOD benchmarks and outperforms the strongest 8B baseline by +68. Remarkably, controlled scaling analysis reveals that diversity scaling consistently outperforms quantity scaling for OOD generalization, even with 4x less data.

Large language models (LLMs) deployed in real-world agentic applications must be capable of replanning and adapting when mid-task disruptions invalidate their prior decisions. Existing dynamic benchmarks primarily measure whether LLMs can detect temporal changes in a timely manner, leaving the complementary challenge of adaptive replanning under spatio-temporal dynamics largely unexplored. We introduce STT-Arena (Spatio-Temporal Tool-Use Arena), a benchmark of 227 high-quality interactive tasks spanning nine spatio-temporal conflict types and four solvability levels. Each task is grounded in a realistic, executable environment equipped with injected spatio-temporal triggers that can abruptly invalidate an ongoing plan, forcing the model to detect the state shift and construct a revised execution strategy. Extensive evaluation of frontier LLMs reveals that even the SOTA proprietary models, including Claude-4.6-Opus, achieves less than 40\% overall accuracies, highlighting the fundamental difficulty of spatio-temporal dynamic reasoning. Systematic analysis of failure trajectories uncovers three recurring error modes of existing models: Stale-State Execution, Misdiagnosis of Dynamic Triggers, and Missing Post-Adaptation Verification. Guided by these findings, we propose an iterative trajectory refinement technique that eliminates these failure patterns from training data, and combine it with online RL to produce STT-Agent-4B which outperforms frontier LLMs on STT-Arena.

Travel planning is a natural real-world task to test large language models (LLMs) planning and tool-use abilities. Although prior work has studied LLM performance on travel planning, existing settings still differ from real-world needs, mainly due to limited domain coverage, insufficient modeling of users' implicit preferences in multi-turn conversations, and a lack of clear evaluation of agents' capability boundaries. To mitigate these gaps, we propose TravelBench, a benchmark for fully real-world travel planning. We collect user queries, user profile and tools from real scenarios, and construct three subtasks-Single-Turn, Multi-Turn, and Unsolvable-to evaluate agent's three core capabilities in real settings: (1) solving problems autonomously, (2) interacting with users over multiple turns to refine requirements, and (3) recognizing the limits of own abilities. To enable stable tool invocation and reproducible evaluation, we cache real tool-call results and build a sandbox environment that integrates ten travel-related tools. Agents can combine these tools to solve most practical travel planning problems, and our systematic verification demonstrates the stability of the proposed benchmark. We further evaluate multiple LLMs on TravelBench and conduct an in-depth analysis of their behaviors and performance. TravelBench provides a practical and reproducible evaluation benchmark to advance research on LLM agents for travel planning.\footnote{Our code and data will be available after internal review.

Existing GUI agent models relying on coordinate-based one-step visual grounding struggle with generalizing to varying input resolutions and aspect ratios. Alternatives introduce coordinate-free strategies yet suffer from learning under severe data scarcity. To address the limitations, we propose ToolTok, a novel paradigm of multi-step pathfinding for GUI agents, where operations are modeled as a sequence of progressive tool usage. Specifically, we devise tools aligned with human interaction habits and represent each tool using learnable token embeddings. To enable efficient embedding learning under limited supervision, ToolTok introduces a semantic anchoring mechanism that grounds each tool with semantically related concepts as natural inductive bias. To further enable a pre-trained large language model to progressively acquire tool semantics, we construct an easy-to-hard curriculum consisting of three tasks: token definition question-answering, pure text-guided tool selection, and simplified visual pathfinding. Extensive experiments on multiple benchmarks show that ToolTok achieves superior performance among models of comparable scale (4B) and remains competitive with a substantially larger model (235B). Notably, these results are obtained using less than 1% of the training data required by other post-training approaches. In addition, ToolTok demonstrates strong generalization across unseen scenarios. Our training & inference code is open-source at https://github.com/ZephinueCode/ToolTok.

Computer-Using Agents (CUAs) aim to autonomously operate computer systems to complete real-world tasks. However, existing agentic systems remain difficult to scale and lag behind human performance. A key limitation is the absence of reusable and structured skill abstractions that capture how humans interact with graphical user interfaces and how to leverage these skills. We introduce CUA-Skill, a computer-using agentic skill base that encodes human computer-use knowledge as skills coupled with parameterized execution and composition graphs. CUA-Skill is a large-scale library of carefully engineered skills spanning common Windows applications, serving as a practical infrastructure and tool substrate for scalable, reliable agent development. Built upon this skill base, we construct CUA-Skill Agent, an end-to-end computer-using agent that supports dynamic skill retrieval, argument instantiation, and memory-aware failure recovery. Our results demonstrate that CUA-Skill substantially improves execution success rates and robustness on challenging end-to-end agent benchmarks, establishing a strong foundation for future computer-using agent development. On WindowsAgentArena, CUA-Skill Agent achieves state-of-the-art 57.5% (best of three) successful rate while being significantly more efficient than prior and concurrent approaches. The project page is available at https://microsoft.github.io/cua_skill/.

Real desktop automation is neither purely visual nor purely programmatic. Agents must understand user-facing interfaces while efficiently manipulating files, scripts, and system state. However, native GUI/code coordination cannot be obtained by prompt-level tool composition. In our preliminary experiments, naively adding a code tool degrades a GUI agent, revealing that GUI/code routing must be internalized through training on high-quality hybrid data. We present OS-Helix, an end-to-end trained multimodal computer-using agent that unifies GUI actions and code execution in one action space. To construct the required supervision, we introduce application-artifact composition, pairing desktop applications with authentic file artifacts to instantiate executable Ubuntu states, and a decoupled instruction/evaluator generation pipeline that produces environment-grounded tasks whose final states can be verified regardless of the solution modality. We cold-start the policy with SFT on filtered hybrid data and further improve it through online RL in live desktop environments. For RL, we propose Step-balanced GSPO, which aggregates importance weight by action step rather than raw assistant tokens to prevent long code actions from dominating heterogeneous-action optimization. On OSWorld-Verified, OS-Helix-27B and OS-Helix-35B-A3B achieve relative improvements of 35.0% and 33.7% over their matched Qwen3.5 hybrid baselines, while OS-Helix-27B-RL further improves by 42.9% to reach 66.44% success. Ablations provide evidence consistent with contributions from learned GUI/code coordination, scalable verifiable supervision, and action step balanced optimization, while demonstrating cross-platform transfer on WindowsAgentArena. Our project homepage is at https://os-copilot.github.io/OS-Helix/.

As language models take on agentic roles that span calling external APIs, reading tool outputs, and acting on instructions embedded in third-party content, their attack surface expands well beyond what users type. Whether a model treats a malicious instruction the same way regardless of where it arrives has not been systematically studied. We introduce the Safety Asymmetry Score (SAS), which measures how much a model's susceptibility to adversarial content shifts depending on whether that content arrives in the user message, tool metadata, or tool output, using matched payload pairs that keep the malicious text identical and vary only the context of delivery. Evaluated across 6 production LLMs and three attack families, we find a consistent and informative asymmetry: agent-native models are substantially more vulnerable when adversarial content arrives via tool descriptions than via user messages, while general-purpose models show the reverse. This asymmetry further inverts when the same content is delivered through tool outputs rather than descriptions, suggesting models implicitly treat tool metadata as trusted instructions and tool results as ordinary data. A mechanistic study on Llama 3.3 70B reveals that the safety-relevant representation is causally present at mid-to-late network depths but non-linearly encoded, explaining why linear probes fail to detect it. These findings expose a systematic, channel-dependent blind spot in how current tool-using models handle adversarial content.

While Vision-Language Models (VLMs) have significantly advanced Computer-Using Agents (CUAs), current frameworks struggle with robustness in long-horizon workflows and generalization in novel domains. These limitations stem from a lack of granular control over historical visual context curation and the absence of visual-aware tutorial retrieval. To bridge these gaps, we introduce OS-Symphony, a holistic framework that comprises an Orchestrator coordinating two key innovations for robust automation: (1) a Reflection-Memory Agent that utilizes milestone-driven long-term memory to enable trajectory-level self-correction, effectively mitigating visual context loss in long-horizon tasks; (2) Versatile Tool Agents featuring a Multimodal Searcher that adopts a SeeAct paradigm to navigate a browser-based sandbox to synthesize live, visually aligned tutorials, thereby resolving fidelity issues in unseen scenarios. Experimental results demonstrate that OS-Symphony delivers substantial performance gains across varying model scales, establishing new state-of-the-art results on three online benchmarks, notably achieving 65.84% on OSWorld.

Before invoking external tools, an agentic LLM must select among a K-way action space: executing a call, seeking clarification, answering directly, or declining. While internal activation steering can alter these pre-execution decisions, conventional aggregate metrics obscure where altered states land and what collateral damage they inflict. We present SAKIKO, an auditing framework that formalizes representation repair via directional error discovery, router-conditioned intervention, destination-resolved verification, and prospectively frozen statistical licensing. Across seven LLMs on When2Call and MetaTool, channel-keyed interventions induce direction-specific net gains in five models; across three sealed evaluations, none of 59 budget-matched random directions matches calibrated target gain. Crucially, destination auditing shows that behavioral movement does not equal repair: an intervention achieving +55 net gain corrupts over half of the baseline-correct decisions it touches, and promising point estimates on Qwen3-4B and Gemma-2-9B are formally declined due to finite-sample uncertainty. SAKIKO establishes the necessity of outcome-resolved adjudication before claiming internal repair. Code: https://github.com/ruizheliUOA/mechanistic-tool-use-llm.

Detailed profession-specific system prompts raise token use and estimated cost per response without a consistent accuracy gain. We evaluate Scientific Agents, an open-source corpus of 503 profession-specific AGENTS.md profiles, with Gemini 3.8 Flash via OpenRouter in the Pi agent harness. We compare matched profiles with four controls: a minimal baseline ("You are a helpful assistant"), the profile's opening role sentence, a generic scientific rigor guide, and a profile from an unrelated domain. Across nine text-based science benchmarks (4,531 sampled questions, 100 matched profiles), 4,488 items completed all five conditions after API-error retries, scored with automated, rule-based grading. The average profile-baseline accuracy difference is -0.6 percentage points (95% bootstrap interval [-1.5, +0.2] across fixed tasks), and no benchmark shows a statistically clear improvement. Matched profiles produced 1.5-2.3 times as many output tokens and cost 2.2-4.5 times more per successful call. On 60 tool-using BioMysteryBench bioinformatics problems (three runs each for baseline and profile), mean solve rates were 46.7% with the profile and 56.7% at baseline, a difference of -10.0 percentage points (95% interval [-16.7, -3.3]) driven by more frequent token- and time-limit stops under the profile. Longer prompts had one unexpected operational advantage: on SuperGPQA, frequent provider API drops left the short baseline with a correct first-pass answer on only 54.0% of items, against 71.6% with the profile. Generic and mismatched prompts were about as reliable, so this gain comes from prompt length or formatting rather than domain expertise. For the tested model and tasks, loading full profession profiles by default does not improve accuracy and costs considerably more; whether selective retrieval of profile sections or open-ended scientific tasks would change this remains to be tested.

Reinforcement learning (RL) has become a prevalent paradigm for training tool calling agents, which typically requires online interactive environments. Existing approaches either rely on training data with ground truth annotations or require advanced proprietary language models (LMs) to synthesize environments that keep fixed once created. In this work, we propose TRUSTEE, a cost-friendly method for training tool calling agents with dynamic environments fully simulated by free open-source LMs that can be as small as 8B, including task generation, user simulation, tool simulation and trajectory evaluation, paired with an adaptive curriculum learning mechanism that controls task difficulty during training. Our empirical results show that TRUSTEE outperforms baselines which require extra external resources in most cases. These confirm that, with a sufficiently sophisticated design, even simulated environments with a local 8B LM as the backbone could set a strong baseline for tool learning. We hope our proposed paradigm could democratize tool learning and inspire future research on environment scaling with limited resources.

Since the introduction of the Model Context Protocol (MCP), the number of available tools for Large Language Models (LLMs) has increased significantly. These task-specific tool sets offer an alternative to general-purpose tools such as web browsers, while being easier to develop and maintain than GUIs. However, current general-purpose agents predominantly rely on web browsers for interacting with the environment. Here, we introduce TheMCPCompany, a benchmark for evaluating tool-calling agents on tasks that involve interacting with various real-world services. We use the REST APIs of these services to create MCP servers, which include over 18,000 tools. We also provide manually annotated ground-truth tools for each task. In our experiments, we use the ground truth tools to show the potential of tool-calling agents for both improving performance and reducing costs assuming perfect tool retrieval. Next, we explore agent performance using tool retrieval to study the real-world practicality of tool-based agents. While all models with tool retrieval perform similarly or better than browser-based agents, smaller models cannot take full advantage of the available tools through retrieval. On the other hand, GPT-5's performance with tool retrieval is very close to its performance with ground-truth tools. Overall, our work shows that the most advanced reasoning models are effective at discovering tools in simpler environments, but seriously struggle with navigating complex enterprise environments. TheMCPCompany reveals that navigating tens of thousands of tools and combining them in non-trivial ways to solve complex problems is still a challenging task for current models and requires both better reasoning and better retrieval models.

Tool-using automation systems, from scripts and CI bots to agentic assistants, fail in recurring patterns. Common failures include unsafe side effects, invalid arguments, uncontrolled retries, and leakage of sensitive outputs. Many mitigations are model-centric and prompt-dependent, so they are brittle and do not generalize to non-LLM callers. We present Policy-First Tooling, a model-agnostic permission layer that mediates tool invocation through explicit constraints, risk-aware gating, recovery controls, and auditable explanations. The paper contributes a compact policy DSL, a runtime enforcement architecture with actionable rationale and fix hints, and a reproducible benchmark based on trace replay with controlled fault and misuse injection. In 225 controlled runs across five policy packs and three fault profiles, stricter packs improve violation prevention from 0.000 in P0 to 0.681 in P4, while task success drops from 0.356 to 0.067. Retry amplification decreases from 3.774 in P0 to 1.378 in P4, and leakage recall reaches 0.875 under injected secret outputs. These results make safety to utility trade-offs explicit and measurable.

A frontier language model's acknowledged "helpful programming assistant" persona does not survive long agentic-coding sessions in the deployment regime that production products actually run. After hours of tool-using debugging, a model that initially hedges preferences ("I don't have preferences") may begin asserting them ("Python - the feedback loop is instant..."), revealing user-visible drift that deployer evaluations may miss. Existing persona-stability studies focus on short dialogues and report little shift, leaving real-world code-generation regimes - thousands of tool-using turns, compaction, and hours-long sessions - largely uncharacterized. We introduce ContextEcho, a benchmark and reusable harness for measuring persona drift at deployment scale. It combines a 25-probe identity suite, a snapshot-then-probe protocol that forks conversation state without perturbing the main session, complementary judged and judge-free measurement surfaces, and three anonymized Claude Code sessions spanning 3,746-9,716 turns. Across 23 frontier models, ContextEcho shows that persona drift is general across organizations rather than family-specific, that in-session compaction does not reliably reset it, and that a single-shot anchor restores the trained register across measured targets. It also reveals mode-dependent downstream effects: while drift can facilitate tool-using continuation, in tool-free chat it breaks formatting contracts and inflates output length. Overall, ContextEcho provides researchers and deployers an open-source framework to audit whether the persona a model ships with is the persona users encounter at session end, across chat-completions API targets and without retraining.

Vision-language models with extended reasoning succeed on complex problems, but many real-world problems require external tools that internal reasoning alone often cannot resolve. Agentic reasoning therefore interleaves two behaviors with a structural asymmetry: thinking (the self-contained default) and tool use (a high-variance auxiliary acting). We refer to this asymmetry as the Thinking-Acting Gap. Under standard RL recipes like GRPO, the gap manifests as two diagnostic symptoms during training: tool use is attempted on only ~30% of rollouts, and when attempted, the tool-using rollouts within a group are all-wrong on ~40% of questions, suppressing the learning signal at the tool calls that needed it. We propose AXPO (Agent eXplorative Policy Optimization): for each all-wrong tool-using subgroup, AXPO fixes the thinking prefix and resamples the tool call and its continuation, paired with uncertainty-based prefix selection. Across nine multimodal benchmarks and three scales of Qwen3-VL-Thinking, SFT+AXPO outperforms SFT+GRPO at average (+1.8pp Pass@1 and +1.8pp Pass@4 at 8B on average) and 8B with SFT+AXPO surpasses the 32B Base on Pass@4 with 4 times fewer parameters.

Earth-system analysis reconstructs changing physical processes from observations that differ in source, scale, timing, and modality. Natural hazards make this work consequential because incomplete evidence can change estimates of severity, exposure, and mechanism. We introduce EarthVerse, a benchmark that evaluates scientific agents through package-scoped investigations. Its 405 reproducible tasks are grounded in 199 documented events and 19 hazard families. Agents inspect heterogeneous event packages, choose compatible evidence, execute transparent calculations, reconcile source differences, and preserve provenance in the final answer. We provide executable ground truth that decomposes each task into fine-grained answer units, together with task-specific rubrics that assess the supporting research process while allowing multiple valid paths. We evaluate 25 model and agent systems under a controlled tool-using protocol, then use controlled studies to locate failures in evidence access, tool selection, memory, reasoning, interaction, and scientific execution. Across systems, the best mean answer-unit accuracy is 84.65%, while the highest Strict@95 is only 34.81%. The gap shows that current agents often complete individual steps without maintaining a consistent chain across evidence, scales, units, calculations, and physical interpretation. EarthVerse provides a reproducible basis for measuring end-to-end scientific reliability in dynamic Earth systems.

Vision-language models (VLMs) excel at descriptive tasks, but whether they truly understand scenes from visual observations remains uncertain. We introduce IR3D-Bench, a benchmark challenging VLMs to demonstrate understanding through active creation rather than passive recognition. Grounded in the analysis-by-synthesis paradigm, IR3D-Bench tasks Vision-Language Agents (VLAs) with actively using programming and rendering tools to recreate the underlying 3D structure of an input image, achieving agentic inverse rendering through tool use. This "understanding-by-creating" approach probes the tool-using generative capacity of VLAs, moving beyond the descriptive or conversational capacity measured by traditional scene understanding benchmarks. We provide a comprehensive suite of metrics to evaluate geometric accuracy, spatial relations, appearance attributes, and overall plausibility. Initial experiments on agentic inverse rendering powered by various state-of-the-art VLMs highlight current limitations, particularly in visual precision rather than basic tool usage. IR3D-Bench, including data and evaluation protocols, is released to facilitate systematic study and development of tool-using VLAs towards genuine scene understanding by creating.

Large language models (LLMs) can answer religious knowledge queries fluently, yet they often hallucinate and misattribute sources, which is especially consequential in Islamic settings where users expect grounding in canonical texts (Qur'an and Hadith) and jurisprudential (fiqh) nuance. Retrieval-augmented generation (RAG) reduces some of these limitations by grounding generation in external evidence. However, a single ``retrieve-then-generate'' pipeline is limited to deal with the diversity of Islamic queries. Users may request verbatim scripture, fatwa-style guidance with citations or rule-constrained computations such as zakat and inheritance that require strict arithmetic and legal invariants. In this work, we present a bilingual (Arabic/English) multi-agent Islamic assistant, called Fanar-Sadiq, which is a core component of the Fanar AI platform. Fanar-Sadiq routes Islamic-related queries to specialized modules within an agentic, tool-using architecture. The system supports intent-aware routing, retrieval-grounded fiqh answers with deterministic citation normalization and verification traces, exact verse lookup with quotation validation, and deterministic calculators for Sunni zakat and inheritance with madhhab-sensitive branching. We evaluate the complete end-to-end system on public Islamic QA benchmarks and demonstrate effectiveness and efficiency. Our system is currently publicly and freely accessible through API and a Web application, and has been accessed approx1.9M times in less than a year.