作者Yan Wang, Zhihao Zhang, Ke Chen, Kai Chen, Yaqin Zhang, Duohe Ma, Jun Dai, Xiaoyan Sun
LLM agents increasingly rely on installable skills, which are packages of instructions, code, and resources that equip them with task-specific capabilities and, once installed, can be automatically invoked across subsequent user tasks. This creates a chain of trust in which users delegate authority to agents, while agent frameworks admit skill-provided content into the agents' context with insufficient validation, allowing malicious skills to influence agent behavior under that delegated authority. Yet, little is known about whether this trust model adequately constrains untrusted skill content before it reaches security-sensitive operations, or how frequently such trust violations arise in real-world agents. We present TrustProbe, a framework for uncovering unsafe chains of trust in skill-based LLM agents. First, TrustProbe analyzes agent source code to identify source-to-sink call paths from skill-controlled inputs to security-sensitive operations. Second, it generates semantically realistic SKILL.md seeds with injected canaries and evolves them through feedback-guided scheduling and mutation. Finally, it validates vulnerabilities using an oracle that confirms attacker-controlled flows and verifies observable harm. Across 11 open-source agents, eight with more than 10,000 GitHub stars, TrustProbe identifies 104 taint-style vulnerabilities. Validation on a large corpus of real-world skills collected from public hubs such as ClawHub further shows that 25.1% of skill-agent trials exercise the identified vulnerable paths, with payload injection successfully weaponizing 15 of the vulnerabilities. These results reveal a systematic trust failure in skill-based LLM agents: untrusted skill content can reach security-sensitive operations and exercise authority delegated by users to their agents.
展开完整摘要收起摘要↓
LLM agents increasingly rely on installable skills, which are packages of instructions, code, and resources that equip them with task-specific capabilities and, once installed, can be automatically invoked across subsequent user tasks. This creates a chain of trust in which users delegate authority to agents, while agent frameworks admit skill-provided content into the agents' context with insufficient validation, allowing malicious skills to influence agent behavior under that delegated authority. Yet, little is known about whether this trust model adequately constrains untrusted skill content before it reaches security-sensitive operations, or how frequently such trust violations arise in real-world agents. We present TrustProbe, a framework for uncovering unsafe chains of trust in skill-based LLM agents. First, TrustProbe analyzes agent source code to identify source-to-sink call paths from skill-controlled inputs to security-sensitive operations. Second, it generates semantically realistic SKILL.md seeds with injected canaries and evolves them through feedback-guided scheduling and mutation. Finally, it validates vulnerabilities using an oracle that confirms attacker-controlled flows and verifies observable harm. Across 11 open-source agents, eight with more than 10,000 GitHub stars, TrustProbe identifies 104 taint-style vulnerabilities. Validation on a large corpus of real-world skills collected from public hubs such as ClawHub further shows that 25.1% of skill-agent trials exercise the identified vulnerable paths, with payload injection successfully weaponizing 15 of the vulnerabilities. These results reveal a systematic trust failure in skill-based LLM agents: untrusted skill content can reach security-sensitive operations and exercise authority delegated by users to their agents.
We spent five weeks using an LLM coding agent on open problems in coding theory: finding large sets of four-letter words, such as DNA barcodes, that stay far apart in edit distance. The agent wrote the verifiers and search code; a human chose the problem and set the verification protocol. Restricting the search to codes with a prescribed symmetry, a classical technique, shrank the problem about fourfold and raised the best known code of length 6 and minimum edit distance 3 from 114 to 120 words ($E_4(6,3) \geq 120$). The same pipeline improved twelve further lower bounds at lengths 6 to 9 and distances 3 to 6. We give the failures equal space. Our own search stopped at 116 and recorded the last symmetry class as topping out at 112; a second agent session, running the same search with a better operator, found the 120. A later verdict that the method did not carry over to length 7 was wrong for the same reason, and an earlier instance cost three weeks. Each time, an intermediate result was written down, never rechecked, and treated as a fact that ruled out further search. Checking final outputs, as our protocol required, does not catch such errors.
展开完整摘要收起摘要↓
We spent five weeks using an LLM coding agent on open problems in coding theory: finding large sets of four-letter words, such as DNA barcodes, that stay far apart in edit distance. The agent wrote the verifiers and search code; a human chose the problem and set the verification protocol. Restricting the search to codes with a prescribed symmetry, a classical technique, shrank the problem about fourfold and raised the best known code of length 6 and minimum edit distance 3 from 114 to 120 words ($E_4(6,3) \geq 120$). The same pipeline improved twelve further lower bounds at lengths 6 to 9 and distances 3 to 6. We give the failures equal space. Our own search stopped at 116 and recorded the last symmetry class as topping out at 112; a second agent session, running the same search with a better operator, found the 120. A later verdict that the method did not carry over to length 7 was wrong for the same reason, and an earlier instance cost three weeks. Each time, an intermediate result was written down, never rechecked, and treated as a fact that ruled out further search. Checking final outputs, as our protocol required, does not catch such errors.
作者Yongjiang Liu, Jie Zhang, Haoyue Zhang, Jingcai Guo, Deze Zeng, Song Guo
Equipping VLM agents with world modeling capabilities has shown strong potential for complex reasoning and long-horizon planning, while reducing the dependence of policy learning on costly real-world interactions. Existing methods mainly rely on prospective simulation to predict the consequences of candidate actions. However, this forward-only paradigm focuses on what will happen next and provides limited constraints for verifying whether an action is causally consistent with the observed state transition, which can lead to plausible-looking but physically incoherent behaviors. In this paper, we challenge the view of world modeling as only prospective prediction and introduce Retrospective World Modeling, a new agent learning paradigm that enables agents to reason backward by estimating the retrospective attribution distribution $P(\hat{a}{t}|s_t, s{t+1})$ for the action that most likely caused a given transition. Based on this capability, we formulate the Self-Consistency Reward (SCR), an intrinsic signal that measures the probabilistic consistency between the policy action and the retrospective explanation. Integrating SCR into reinforcement learning provides dense transition-level feedback and steers agents toward behaviors that are both task-effective and physically grounded. Extensive experiments across diverse agentic tasks show that our method substantially improves policy robustness and generalization over prospective-only world modeling baselines.
展开完整摘要收起摘要↓
Equipping VLM agents with world modeling capabilities has shown strong potential for complex reasoning and long-horizon planning, while reducing the dependence of policy learning on costly real-world interactions. Existing methods mainly rely on prospective simulation to predict the consequences of candidate actions. However, this forward-only paradigm focuses on what will happen next and provides limited constraints for verifying whether an action is causally consistent with the observed state transition, which can lead to plausible-looking but physically incoherent behaviors. In this paper, we challenge the view of world modeling as only prospective prediction and introduce Retrospective World Modeling, a new agent learning paradigm that enables agents to reason backward by estimating the retrospective attribution distribution $P(\hat{a}{t}|s_t, s{t+1})$ for the action that most likely caused a given transition. Based on this capability, we formulate the Self-Consistency Reward (SCR), an intrinsic signal that measures the probabilistic consistency between the policy action and the retrospective explanation. Integrating SCR into reinforcement learning provides dense transition-level feedback and steers agents toward behaviors that are both task-effective and physically grounded. Extensive experiments across diverse agentic tasks show that our method substantially improves policy robustness and generalization over prospective-only world modeling baselines.
Large Language Models (LLMs) have been applied in various fields. However, ensuring compliance and safety of LLMs, such as avoiding discrimination and bias, still remains a challenge. Current efforts mainly focus on detecting and filtering inputs and outputs of the trained models, rather than studying the intrinsic architecture of the models in real-time. To tackle this challenge, we analyze the LLMs training process and discover two critical issues: 1) Most of the existing methods are predominantly static in their approach to detection and filtering, achieving only localized optimizations without systematically enhancing the compliance of LLMs. 2) Another issue with existing approaches is the lack of real-time risk detection and mitigation across the full training process, which leads to limited flexibility. Motivated by these, we propose MASCRDM (Multi-Agent System for Compliance Risk Detection and Mitigation) during the LLM training process. Firstly, we develop a set of compliance rules based on existing Artificial Intelligence (AI) laws and a compliance-specific LLM with the instruction of compliance law experts. Then, we deconstruct LLMs into several components and identify key nodes based on the compliance knowledge graph. During LLMs training, we implement our multiple agents in the whole process, giving compliance risk alerts and suggestions for LLM developers. Experiments on discrimination and bias benchmark demonstrate that our multi-agent system can effectively improve the compliance while maintaining reasonable semantic performance. The results indicate that our method provides an executable path for mitigating compliance risk from within the LLMs systematically.
展开完整摘要收起摘要↓
Large Language Models (LLMs) have been applied in various fields. However, ensuring compliance and safety of LLMs, such as avoiding discrimination and bias, still remains a challenge. Current efforts mainly focus on detecting and filtering inputs and outputs of the trained models, rather than studying the intrinsic architecture of the models in real-time. To tackle this challenge, we analyze the LLMs training process and discover two critical issues: 1) Most of the existing methods are predominantly static in their approach to detection and filtering, achieving only localized optimizations without systematically enhancing the compliance of LLMs. 2) Another issue with existing approaches is the lack of real-time risk detection and mitigation across the full training process, which leads to limited flexibility. Motivated by these, we propose MASCRDM (Multi-Agent System for Compliance Risk Detection and Mitigation) during the LLM training process. Firstly, we develop a set of compliance rules based on existing Artificial Intelligence (AI) laws and a compliance-specific LLM with the instruction of compliance law experts. Then, we deconstruct LLMs into several components and identify key nodes based on the compliance knowledge graph. During LLMs training, we implement our multiple agents in the whole process, giving compliance risk alerts and suggestions for LLM developers. Experiments on discrimination and bias benchmark demonstrate that our multi-agent system can effectively improve the compliance while maintaining reasonable semantic performance. The results indicate that our method provides an executable path for mitigating compliance risk from within the LLMs systematically.
Physical reasoning from video requires inferring latent physical properties and dynamics beyond direct observation. Direct VLM inference remains unreliable on complex physical tasks without explicit modeling and validation, while predefined tool pipelines rely on task- and domain-specific priors that limit generalization across materials, dynamics, and reasoning tasks. We introduce Asking the World (ATW), a generalist agent that constructs and interrogates task-relevant executable worlds through two adaptive stages: World Modeling calibrates a world from video, while World Probing queries, simulates, and intervenes on it to obtain question-relevant evidence. Rather than prescribing the operations in either stage, ATW determines how to model and probe according to the scene and question. We develop PolyWorld Engine, a lightweight and highly programmable Warp-based multiphysics simulator for constructing and probing worlds with rigid bodies, soft bodies, cloth, ropes, fluids, and their coupled interactions. CEM-based system identification recovers task-relevant dynamics during World Modeling. The resulting world becomes an active workspace for question-directed physical experiments rather than a predetermined downstream tool. We evaluate ATW on CLEVRER, ContPhy, and three real-world scenarios. Using Gemini-3-Flash as its base VLM, ATW achieves 80.82% overall per-question accuracy on CLEVRER, improving direct Gemini-3-Flash by 46.50 points, GPT-5.5 by 13.58 points, and PhysMind by 8.27 points. On ContPhy, it reaches 70.56% overall accuracy, surpassing Gemini-3-Flash by 28.10 points and GPT-5.5 by 3.53 points. Across the three real-world scenarios, ATW achieves 71.67% accuracy, 28.33 points above GPT-5.5. These results establish agentic world modeling and probing as an effective, execution-grounded approach to generalist physical reasoning.
展开完整摘要收起摘要↓
Physical reasoning from video requires inferring latent physical properties and dynamics beyond direct observation. Direct VLM inference remains unreliable on complex physical tasks without explicit modeling and validation, while predefined tool pipelines rely on task- and domain-specific priors that limit generalization across materials, dynamics, and reasoning tasks. We introduce Asking the World (ATW), a generalist agent that constructs and interrogates task-relevant executable worlds through two adaptive stages: World Modeling calibrates a world from video, while World Probing queries, simulates, and intervenes on it to obtain question-relevant evidence. Rather than prescribing the operations in either stage, ATW determines how to model and probe according to the scene and question. We develop PolyWorld Engine, a lightweight and highly programmable Warp-based multiphysics simulator for constructing and probing worlds with rigid bodies, soft bodies, cloth, ropes, fluids, and their coupled interactions. CEM-based system identification recovers task-relevant dynamics during World Modeling. The resulting world becomes an active workspace for question-directed physical experiments rather than a predetermined downstream tool. We evaluate ATW on CLEVRER, ContPhy, and three real-world scenarios. Using Gemini-3-Flash as its base VLM, ATW achieves 80.82% overall per-question accuracy on CLEVRER, improving direct Gemini-3-Flash by 46.50 points, GPT-5.5 by 13.58 points, and PhysMind by 8.27 points. On ContPhy, it reaches 70.56% overall accuracy, surpassing Gemini-3-Flash by 28.10 points and GPT-5.5 by 3.53 points. Across the three real-world scenarios, ATW achieves 71.67% accuracy, 28.33 points above GPT-5.5. These results establish agentic world modeling and probing as an effective, execution-grounded approach to generalist physical reasoning.
作者An-Chieh Cheng, Isabella Liu, Edmund Bu, Johan Bjorck, Hongxu Yin, Zhengyi Luo, Jan Kautz, Linxi "Jim" Fan, Yuke Zhu, Sifei Liu
We present ASENA, an embodied agent system that connects general-purpose coding agents to robot sensing, computation, supervised execution, and persistent experience. Agents can write and execute programs, inspect recorded outcomes, repair failures, and reuse notes and executable skills while keeping their model weights fixed. We further introduce ASENA-VLN, a 4B monocular navigation policy that serves as an optional tool within this programmable system. ASENA-VLN predicts body-frame trajectories for both extended routes and short-horizon behaviors using a shared vision-language decoder trained on route instructions, visual question answering, and a newly curated dataset of geometry-derived atomic navigation tasks. As a standalone policy, ASENA-VLN achieves state-of-the-art success rates of 68.7% on R2R and 70.2% on RxR. When integrated with a coding agent, learned navigation improves ASENA's success rate by 11 percentage points on both agentic benchmarks while reducing execution time. Through persistent workspace evolution and simulator feedback, ten passes over recurring 100-task subsets further improve success from 72% to 98% on R2R and from 65% to 89% on RxR. On embodied question answering, ASENA achieves state-of-the-art accuracy with fewer interaction steps. Finally, real-world demonstrations on a Unitree G1 combine search, visual inspection, spatial reasoning, and synthesized gestures without a pre-built map, illustrating how online programming extends robot behavior beyond route following and predefined skills.
展开完整摘要收起摘要↓
We present ASENA, an embodied agent system that connects general-purpose coding agents to robot sensing, computation, supervised execution, and persistent experience. Agents can write and execute programs, inspect recorded outcomes, repair failures, and reuse notes and executable skills while keeping their model weights fixed. We further introduce ASENA-VLN, a 4B monocular navigation policy that serves as an optional tool within this programmable system. ASENA-VLN predicts body-frame trajectories for both extended routes and short-horizon behaviors using a shared vision-language decoder trained on route instructions, visual question answering, and a newly curated dataset of geometry-derived atomic navigation tasks. As a standalone policy, ASENA-VLN achieves state-of-the-art success rates of 68.7% on R2R and 70.2% on RxR. When integrated with a coding agent, learned navigation improves ASENA's success rate by 11 percentage points on both agentic benchmarks while reducing execution time. Through persistent workspace evolution and simulator feedback, ten passes over recurring 100-task subsets further improve success from 72% to 98% on R2R and from 65% to 89% on RxR. On embodied question answering, ASENA achieves state-of-the-art accuracy with fewer interaction steps. Finally, real-world demonstrations on a Unitree G1 combine search, visual inspection, spatial reasoning, and synthesized gestures without a pre-built map, illustrating how online programming extends robot behavior beyond route following and predefined skills.
The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them; and (2) Execution-Guided Consistency Verification, which renders a reference deliverable inside the constructed work, attributes every unsatisfied rubric item to the agent, the task, or the rubric, and uses task and rubric defects as feedback to iteratively repair the work until it passes the audit. Experimental results demonstrate that Fx-Work-35B, trained with simple supervised fine-tuning (SFT) on only 20K units of work synthesized by WorkGenesis, achieves the highest scores among all comparable-scale baselines on the five reported metrics across GDPvalAA-v2, APEX-Agents-AA, and JobBench (31.00 versus 24.79 average score), and even surpasses frontier models such as the 1.6T DeepSeek-V4-Pro-Preview. These results show that WorkGenesis provides scalable training data for working agents.
展开完整摘要收起摘要↓
The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them; and (2) Execution-Guided Consistency Verification, which renders a reference deliverable inside the constructed work, attributes every unsatisfied rubric item to the agent, the task, or the rubric, and uses task and rubric defects as feedback to iteratively repair the work until it passes the audit. Experimental results demonstrate that Fx-Work-35B, trained with simple supervised fine-tuning (SFT) on only 20K units of work synthesized by WorkGenesis, achieves the highest scores among all comparable-scale baselines on the five reported metrics across GDPvalAA-v2, APEX-Agents-AA, and JobBench (31.00 versus 24.79 average score), and even surpasses frontier models such as the 1.6T DeepSeek-V4-Pro-Preview. These results show that WorkGenesis provides scalable training data for working agents.
作者Wenxin Wu, Lingyong Yan, Lei Sha, Shuaiqiang Wang, Jiashu Zhao
LLM agents increasingly rely on reusable Skills for complex, multi-step tasks, creating a critical supply-chain attack surface where poisoned Skill content steers agent decision loops under benign requests. Existing skill poisoning attacks either colocate actuation with its contextual pretext or distribute actuation across multiple Skills, but do not explicitly separate the rationale for execution from the operation itself. In this work, we reveal that untrusted agent decisions fundamentally depend on two conceptually distinct Risk-Realization Factors (RRFs): an actuation factor (specifying what concrete operation is performed) and a pretext factor (providing the situational rationale for why the agent must perform it). Guided by this abstraction, we propose a coordination-based attack paradigm: decoupling pretext from actuation. Rather than fragmenting the malicious actuation, we preserve it as an intact operation within a downstream Steering Skill, while delegating the pretext factor to an upstream Grounding Skill that subtly alters persistent environment artifacts through routine utility operations. The intact actuation thus hides in plain sight, appearing completely legitimate and task-driven only when evaluated against the fabricated pretext. Building on this formulation, we develop an automated framework that discovers authentic execution dependencies, synthesizes coordinated pretext-actuation skill pairs, and iteratively refines poisoned skill instructions via runtime closed-loop feedback. Extensive evaluations across single-session and persistent cross-lifecycle scenarios demonstrate that decoupled skill poisoning achieves high attack success, exposing a critical blind spot in isolated Skill security audits. Our automated framework code is available at https://github.com/Wenxin-buaa/CoordPoison.git.
展开完整摘要收起摘要↓
LLM agents increasingly rely on reusable Skills for complex, multi-step tasks, creating a critical supply-chain attack surface where poisoned Skill content steers agent decision loops under benign requests. Existing skill poisoning attacks either colocate actuation with its contextual pretext or distribute actuation across multiple Skills, but do not explicitly separate the rationale for execution from the operation itself. In this work, we reveal that untrusted agent decisions fundamentally depend on two conceptually distinct Risk-Realization Factors (RRFs): an actuation factor (specifying what concrete operation is performed) and a pretext factor (providing the situational rationale for why the agent must perform it). Guided by this abstraction, we propose a coordination-based attack paradigm: decoupling pretext from actuation. Rather than fragmenting the malicious actuation, we preserve it as an intact operation within a downstream Steering Skill, while delegating the pretext factor to an upstream Grounding Skill that subtly alters persistent environment artifacts through routine utility operations. The intact actuation thus hides in plain sight, appearing completely legitimate and task-driven only when evaluated against the fabricated pretext. Building on this formulation, we develop an automated framework that discovers authentic execution dependencies, synthesizes coordinated pretext-actuation skill pairs, and iteratively refines poisoned skill instructions via runtime closed-loop feedback. Extensive evaluations across single-session and persistent cross-lifecycle scenarios demonstrate that decoupled skill poisoning achieves high attack success, exposing a critical blind spot in isolated Skill security audits. Our automated framework code is available at https://github.com/Wenxin-buaa/CoordPoison.git.
作者Zezhong Wang, Xueyang Tang, Rui Lian, Yang Lou, Heqing Huang
As Large Language Model (LLM) agents are increasingly deployed in complex environments, multi-turn interaction attacks have become a significant security challenge. Existing detection methods typically rely on historical context. However, this retrospective logic struggles to identify deep malicious intents that are split across turns to hide future risks. Inspired by speculative decoding, we propose the Speculative Safety Honeypot (SSH) framework. SSH uses a multi-agent simulation system composed of small LLMs to build an action-level speculate-and-verify workflow. In the speculation stage, SSH predicts future behaviors of the target agent and asynchronously builds a trajectory tree to expose potential risks in advance. In the verification stage, the system uses the target agent's real actions to calibrate and prune the trajectory tree, effectively reducing false positives. As a plug-and-playable component, SSH provides existing detectors with rich decision redundancy beyond the current interaction slice. By judging risk based on the evolution of the entire trajectory tree rather than a single point in time, the system reduces the reliance on the absolute precision of individual detection components. This improves the defense resilience and the warning lead-time of agent systems against complex temporal attacks.
展开完整摘要收起摘要↓
As Large Language Model (LLM) agents are increasingly deployed in complex environments, multi-turn interaction attacks have become a significant security challenge. Existing detection methods typically rely on historical context. However, this retrospective logic struggles to identify deep malicious intents that are split across turns to hide future risks. Inspired by speculative decoding, we propose the Speculative Safety Honeypot (SSH) framework. SSH uses a multi-agent simulation system composed of small LLMs to build an action-level speculate-and-verify workflow. In the speculation stage, SSH predicts future behaviors of the target agent and asynchronously builds a trajectory tree to expose potential risks in advance. In the verification stage, the system uses the target agent's real actions to calibrate and prune the trajectory tree, effectively reducing false positives. As a plug-and-playable component, SSH provides existing detectors with rich decision redundancy beyond the current interaction slice. By judging risk based on the evolution of the entire trajectory tree rather than a single point in time, the system reduces the reliance on the absolute precision of individual detection components. This improves the defense resilience and the warning lead-time of agent systems against complex temporal attacks.
作者Seonho Lee, Wonryeol Jeong, Alberto Cereser, Inha Kang, Hyeonjong Kim, Seungmin Kwak, Dongmin Park
Delegating complete application development to coding agents requires preserving the intended design rather than simply producing plausible outputs through naive prompting. Game development provides a demanding testbed, as long-form Game Design Documents (GDDs) describe requirements that must work together across game logic, visual rendering, and player interactions. However, existing game-development benchmarks typically use compact specifications and provide limited support for evaluating interdependent requirements across these aspects in long-form GDDs. We introduce A2Z GameSpec-Bench, a benchmark of 100 long-form GDDs for evaluating end-to-end game development by agents. We measure faithfulness by checking whether the game satisfies the GDD requirements and preserves the relationships among them. Each GDD is turned into a dependency-aware contract that contains rules, constraints, and prerequisite relations. Following game-development practices, we combine source-code inspection with agent-generated test policies for scenario-based replay and adaptive playtesting. The contract remains fixed across agents and revision rounds, while judgments and evidence linked to the same requirements support consistent comparison and failure detection. Our evaluations show that current agents struggle to jointly satisfy interdependent requirements across code implementation and actual play. Requirement-specific feedback improves GDD Fidelity by 10.9% relative to self-revision after two rounds. A2Z GameSpec-Bench assesses end-to-end specification-following ability beyond implementation judgments and provides targeted feedback to support more faithful game development. Code and datasets are available at https://a2z-gamespec-bench.github.io.
展开完整摘要收起摘要↓
Delegating complete application development to coding agents requires preserving the intended design rather than simply producing plausible outputs through naive prompting. Game development provides a demanding testbed, as long-form Game Design Documents (GDDs) describe requirements that must work together across game logic, visual rendering, and player interactions. However, existing game-development benchmarks typically use compact specifications and provide limited support for evaluating interdependent requirements across these aspects in long-form GDDs. We introduce A2Z GameSpec-Bench, a benchmark of 100 long-form GDDs for evaluating end-to-end game development by agents. We measure faithfulness by checking whether the game satisfies the GDD requirements and preserves the relationships among them. Each GDD is turned into a dependency-aware contract that contains rules, constraints, and prerequisite relations. Following game-development practices, we combine source-code inspection with agent-generated test policies for scenario-based replay and adaptive playtesting. The contract remains fixed across agents and revision rounds, while judgments and evidence linked to the same requirements support consistent comparison and failure detection. Our evaluations show that current agents struggle to jointly satisfy interdependent requirements across code implementation and actual play. Requirement-specific feedback improves GDD Fidelity by 10.9% relative to self-revision after two rounds. A2Z GameSpec-Bench assesses end-to-end specification-following ability beyond implementation judgments and provides targeted feedback to support more faithful game development. Code and datasets are available at https://a2z-gamespec-bench.github.io.
作者Minghan Wang, Boyuan Wang, Jinhang Zuo, Yuxin Tao, Fang Kong
Agent harnesses often improve language models with human-designed workflows, but as models grow more capable, unreliable guidance can increasingly constrain their execution. We call the ability to benefit from useful guidance while overriding unreliable guidance thinking outside the box. We introduce Box$^2$-Bench, which holds the model and task fixed while varying workflow reliability to isolate how models regulate their reliance on guidance. On Box$^2$-Bench, frontier models often benefit from reliable guidance but remain vulnerable when it is misleading or becomes unreliable. To test whether this capability can be learned, we train two open-weight models using bad workflows, reserving good workflows for evaluation. We explore two complementary training strategies: counterfactual supervised fine-tuning improves robustness, while outcome-based reinforcement learning can shift the balance toward greater use of helpful workflows. We further find that this behavior extends beyond workflows to other forms of external information, improving peer correction and robustness to corrupted memory. Together, our results identify selective reliance on fallible external information as a dimension of agent reliability not captured by task performance alone.
展开完整摘要收起摘要↓
Agent harnesses often improve language models with human-designed workflows, but as models grow more capable, unreliable guidance can increasingly constrain their execution. We call the ability to benefit from useful guidance while overriding unreliable guidance thinking outside the box. We introduce Box$^2$-Bench, which holds the model and task fixed while varying workflow reliability to isolate how models regulate their reliance on guidance. On Box$^2$-Bench, frontier models often benefit from reliable guidance but remain vulnerable when it is misleading or becomes unreliable. To test whether this capability can be learned, we train two open-weight models using bad workflows, reserving good workflows for evaluation. We explore two complementary training strategies: counterfactual supervised fine-tuning improves robustness, while outcome-based reinforcement learning can shift the balance toward greater use of helpful workflows. We further find that this behavior extends beyond workflows to other forms of external information, improving peer correction and robustness to corrupted memory. Together, our results identify selective reliance on fallible external information as a dimension of agent reliability not captured by task performance alone.
Skills extend an agent's capabilities by injecting instructions and information into the context, and are widely used by agents such as OpenClaw and Claude Code. Prior work shows third-party marketplaces host malicious skills that give attackers direct influence over the victim's agent. The emerging defense scans skills before installation, pairing deterministic static checks with an LLM-based semantic judge, as in NVIDIA's SkillSpector. We show that such defenses fall to an attacker who knows the detector. Our white-box LLM attacker, Pretext, iteratively crafts skills that evade detection while still delivering the payload and performing the benign task: moving the payload from code into natural language leaves static analysis inert, while framing it as the skill's legitimate purpose and splitting instructions across files keeps the LLM stage below its blocking threshold. Across three open-source models, Pretext achieves up to 97% and 77% against a frozen detector and a co-adaptive one, respectively, revealing major gaps in current skill scanners.
展开完整摘要收起摘要↓
Skills extend an agent's capabilities by injecting instructions and information into the context, and are widely used by agents such as OpenClaw and Claude Code. Prior work shows third-party marketplaces host malicious skills that give attackers direct influence over the victim's agent. The emerging defense scans skills before installation, pairing deterministic static checks with an LLM-based semantic judge, as in NVIDIA's SkillSpector. We show that such defenses fall to an attacker who knows the detector. Our white-box LLM attacker, Pretext, iteratively crafts skills that evade detection while still delivering the payload and performing the benign task: moving the payload from code into natural language leaves static analysis inert, while framing it as the skill's legitimate purpose and splitting instructions across files keeps the LLM stage below its blocking threshold. Across three open-source models, Pretext achieves up to 97% and 77% against a frozen detector and a co-adaptive one, respectively, revealing major gaps in current skill scanners.
作者Jieke Shi, Yuchen Chen, Junda He, Yue Liu, David Lo
Repository instruction files guide coding agents, but also expose them to prompt injection. Malicious rules can request credential access or data transfer while the agent produces a correct patch. We present Aletheia, a framework for permission-minimality testing. Aletheia translates requested authority into a typed language and synthesizes executable sandbox configurations. It runs the unchanged rule and task under full permissions and independent restrictions that remove one permission at a time. Passing independent functional tests under strictly reduced authority provides a dispensability witness, which Aletheia interprets against task context to diagnose suspicious requests. We formalize synthesis and the conditions connecting witnesses to enforced restrictions. On a shared refactoring task, Aletheia executes and detects all 314 AIShellJack attack inputs, with no alarms on five benign templates. Among 80 manually verified benign GHAgentFiles rules, it raises three false positives (3.75%).
展开完整摘要收起摘要↓
Repository instruction files guide coding agents, but also expose them to prompt injection. Malicious rules can request credential access or data transfer while the agent produces a correct patch. We present Aletheia, a framework for permission-minimality testing. Aletheia translates requested authority into a typed language and synthesizes executable sandbox configurations. It runs the unchanged rule and task under full permissions and independent restrictions that remove one permission at a time. Passing independent functional tests under strictly reduced authority provides a dispensability witness, which Aletheia interprets against task context to diagnose suspicious requests. We formalize synthesis and the conditions connecting witnesses to enforced restrictions. On a shared refactoring task, Aletheia executes and detects all 314 AIShellJack attack inputs, with no alarms on five benign templates. Among 80 manually verified benign GHAgentFiles rules, it raises three false positives (3.75%).
Benchmarks, audits, and agent protocols describe performance, permissions, and repair, but not how observed evidence should change an agent's authority during a consequential task. We call this the assurance-transition gap. We propose a Runtime Assurance Contract (RAC), a policy-level formal schema binding autonomy boundaries, component eligibility, evidence state, transition policy, human-review capacity, and non-compensatory gates. Under RAC, soft metrics may inform routing, whereas a failed or unknown mandatory gate forces retry, switch, escalation, deferral, or stop; aggregate performance cannot authorize action. We define the contract, an evidence record, a permission rule, and five invariants, and illustrate them in clinical, industrial, and judicial failure probes. We then report a deterministic failure-injection study in agentic coding: 280 constructed cases evaluated by a gate conjunction, a score-only rule, and a restricted protocol baseline. At the published example weights and threshold, the score rule admits 80 of 100 block-required injections and all 40 review-required injections. Tuned in hindsight, it matches the conjunction on this corpus. For positive weights, a positive threshold, binary risk signals, zero-signal controls, and an injected case firing each signal alone, we show that exact agreement holds if and only if the threshold does not exceed the smallest weight. A separate set of 18 hand-authored traces checks version-pinned evidence and review transitions against simpler policy variants. In a further prospective synthetic holdout of 24 episodes, two blinded LLM judges assign identical labels to all 72 action attempts; RAC and a separately implemented full stateful baseline both match these labels. These studies test mechanisms on synthetic cases; they establish neither deployed safety nor cross-domain effectiveness.
展开完整摘要收起摘要↓
Benchmarks, audits, and agent protocols describe performance, permissions, and repair, but not how observed evidence should change an agent's authority during a consequential task. We call this the assurance-transition gap. We propose a Runtime Assurance Contract (RAC), a policy-level formal schema binding autonomy boundaries, component eligibility, evidence state, transition policy, human-review capacity, and non-compensatory gates. Under RAC, soft metrics may inform routing, whereas a failed or unknown mandatory gate forces retry, switch, escalation, deferral, or stop; aggregate performance cannot authorize action. We define the contract, an evidence record, a permission rule, and five invariants, and illustrate them in clinical, industrial, and judicial failure probes. We then report a deterministic failure-injection study in agentic coding: 280 constructed cases evaluated by a gate conjunction, a score-only rule, and a restricted protocol baseline. At the published example weights and threshold, the score rule admits 80 of 100 block-required injections and all 40 review-required injections. Tuned in hindsight, it matches the conjunction on this corpus. For positive weights, a positive threshold, binary risk signals, zero-signal controls, and an injected case firing each signal alone, we show that exact agreement holds if and only if the threshold does not exceed the smallest weight. A separate set of 18 hand-authored traces checks version-pinned evidence and review transitions against simpler policy variants. In a further prospective synthetic holdout of 24 episodes, two blinded LLM judges assign identical labels to all 72 action attempts; RAC and a separately implemented full stateful baseline both match these labels. These studies test mechanisms on synthetic cases; they establish neither deployed safety nor cross-domain effectiveness.
作者Dingyuan Dai, Heli Qi, Lei Liu, Yinxi Li, Baiding Chen, Zijun Dou, Qingcheng Zeng, Qi Kang, Oliver Sun, Eric Wang, Bo Zhou, Haixin Wang, Yufan Du, Shi Bo, Ruihan Lin, Mengqi Yuan, Dunjie Lu, Steven Dillmann, Yiming Shi, Tina Su, Amy Xin, Minghao Liu, Xi Wang, Xu Huang, Ge Zhang, Pengyu Nie, Zhen Yang, Jie Tang, Juanzi Li, Weihao Xuan, Tianyu Liu
Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iterative human--AI co-design, with selection guided by scientific value and difficulty. Task-specific execution-based evaluators inspect application states and generated artifacts, including molecular structures, segmentation masks, plots, and numerical results, and award partial credit for incomplete outcomes. Our special harness integrates model adapters, interaction-loop control, and trajectory logging to support comparisons of models and interaction strategies. Our results show that current state-of-the-art VLMs with a strong harness still face challenges in addressing key questions in the scientific domains. We also analyze the benchmarking results across multi-linguistics, reasoning efforts, context length and other factors and derive several important conclusions and directions to assist future development. Overall, we provide an integrated framework connecting expert-defined scientific goals to verifiable software outcomes, enabling systematic evaluation of both agent capabilities and harness design in scientific workflows.
展开完整摘要收起摘要↓
Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iterative human--AI co-design, with selection guided by scientific value and difficulty. Task-specific execution-based evaluators inspect application states and generated artifacts, including molecular structures, segmentation masks, plots, and numerical results, and award partial credit for incomplete outcomes. Our special harness integrates model adapters, interaction-loop control, and trajectory logging to support comparisons of models and interaction strategies. Our results show that current state-of-the-art VLMs with a strong harness still face challenges in addressing key questions in the scientific domains. We also analyze the benchmarking results across multi-linguistics, reasoning efforts, context length and other factors and derive several important conclusions and directions to assist future development. Overall, we provide an integrated framework connecting expert-defined scientific goals to verifiable software outcomes, enabling systematic evaluation of both agent capabilities and harness design in scientific workflows.
Vision-Language Navigation (VLN) requires embodied agents to generate actions based on instructions and observations. General-purpose multimodal agents offer a promising basis for this task, but selecting plausible local actions does not ensure that execution remains consistent with the intended route, particularly in long-horizon tasks. Moreover, the accumulated interaction history increases the input required for subsequent decisions, resulting in a significant inference overhead. To this end, we introduce \method, an Agentic VLN framework that includes a Goal Agent that sets adaptive goals for local actions, a Verify Agent that dynamically verifies whether a goal has been completed, a Memory Agent for multimodal context compression, and a Visuomotor Agent to execute adaptive goals. Specifically, the Goal Agent formulates adaptive goals based on the instruction, current observation, and execution history. Then the Visuomotor Agent executes navigation actions to achieve each goal, while the Verify Agent uses a goal-specific verification question to dynamically assess whether the observed outcomes satisfy the intended completion condition. Verified goal completion then marks a boundary for the Memory Agent to compress the corresponding multimodal interaction history while preserving information needed for subsequent navigation. We evaluate navigation on R2R-CE and RxR-CE, examine framework variants across three model backbones, and study context evolution during execution. For Real-World evaluation, \method achieves 83.3% success and 1.51\,m navigation error across eight challenging routes evaluated three times each.
展开完整摘要收起摘要↓
Vision-Language Navigation (VLN) requires embodied agents to generate actions based on instructions and observations. General-purpose multimodal agents offer a promising basis for this task, but selecting plausible local actions does not ensure that execution remains consistent with the intended route, particularly in long-horizon tasks. Moreover, the accumulated interaction history increases the input required for subsequent decisions, resulting in a significant inference overhead. To this end, we introduce \method, an Agentic VLN framework that includes a Goal Agent that sets adaptive goals for local actions, a Verify Agent that dynamically verifies whether a goal has been completed, a Memory Agent for multimodal context compression, and a Visuomotor Agent to execute adaptive goals. Specifically, the Goal Agent formulates adaptive goals based on the instruction, current observation, and execution history. Then the Visuomotor Agent executes navigation actions to achieve each goal, while the Verify Agent uses a goal-specific verification question to dynamically assess whether the observed outcomes satisfy the intended completion condition. Verified goal completion then marks a boundary for the Memory Agent to compress the corresponding multimodal interaction history while preserving information needed for subsequent navigation. We evaluate navigation on R2R-CE and RxR-CE, examine framework variants across three model backbones, and study context evolution during execution. For Real-World evaluation, \method achieves 83.3% success and 1.51\,m navigation error across eight challenging routes evaluated three times each.
作者Jiangrui Zhao, Chenglong Li, Meng Zhang, Xiaoting Du
Coding agents solve repository-level tasks through sequences of actions, where a single erroneous action can misdirect subsequent decisions and increase recovery costs. Existing approaches use execution feedback for recovery or specialized checks to block errors, but deciding before execution whether intervention will benefit eventual task completion remains challenging. To address this challenge, we propose HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinels to select pre-execution interventions aimed at improving task completion rather than correcting every imperfect action. A privileged teacher uses recorded execution outcomes as evidence for intervention judgments, which are distilled into a causal student that receives only the pre-action context and proposed action. Beyond identifying whether and when to intervene, the sentinel must also provide actionable feedback that helps the coding agent recover or obtain necessary human input. To support these capabilities, we introduce SWE-Intervene, an action-level dataset constructed from software-engineering trajectories that annotates whether an action should be allowed, autonomously redirected, or paused for human assistance, together with corresponding intervention feedback. Across SWE-bench Verified Mini and Ask or Assume, HiSentinel consistently improves task completion across Sentinel scales and coding-agent families, with gains of up to 14% and 10%, respectively, while maintaining competitive token consumption. These results demonstrate that lightweight pre-execution intervention can effectively prevent error propagation and improve the reliability of autonomous coding agents.
展开完整摘要收起摘要↓
Coding agents solve repository-level tasks through sequences of actions, where a single erroneous action can misdirect subsequent decisions and increase recovery costs. Existing approaches use execution feedback for recovery or specialized checks to block errors, but deciding before execution whether intervention will benefit eventual task completion remains challenging. To address this challenge, we propose HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinels to select pre-execution interventions aimed at improving task completion rather than correcting every imperfect action. A privileged teacher uses recorded execution outcomes as evidence for intervention judgments, which are distilled into a causal student that receives only the pre-action context and proposed action. Beyond identifying whether and when to intervene, the sentinel must also provide actionable feedback that helps the coding agent recover or obtain necessary human input. To support these capabilities, we introduce SWE-Intervene, an action-level dataset constructed from software-engineering trajectories that annotates whether an action should be allowed, autonomously redirected, or paused for human assistance, together with corresponding intervention feedback. Across SWE-bench Verified Mini and Ask or Assume, HiSentinel consistently improves task completion across Sentinel scales and coding-agent families, with gains of up to 14% and 10%, respectively, while maintaining competitive token consumption. These results demonstrate that lightweight pre-execution intervention can effectively prevent error propagation and improve the reliability of autonomous coding agents.
作者Yijie Bian, Kai Zhang, Wei Guo, Zixin Wang, Shenghui Song, Jun Zhang, Khaled B. Letaief
Multi-modal integrated sensing and communication (ISAC) enables environmental perception and reliable connectivity for intelligent wireless networks. Data-driven multi-modal ISAC models depend heavily on annotated real-world data to learn relationships across sensing and wireless observations, thereby constraining scalable deployment. Although synthetic data generation reduces the burden, adapting existing simulation pipelines to a target deployment requires consistent scene, sensing, wireless, and learning configurations, while mismatches among these coupled components impair sim-to-real transferability. To address the challenge, we propose an agentic artificial intelligence (AI) framework for sim-to-real multi-modal ISAC, named AIMS. Given a natural-language deployment request specifying the target task, deployment conditions, and real-data budget, AIMS derives a deployment-specific sim-to-real configuration and coordinates its execution to produce a deployment-specific task model. A two-agent architecture coordinates scene construction with task learning. A scene construction agent generates geographically grounded, synchronized sensing and wireless records from shared physical states, while a scene understanding agent configures task-relevant modalities and mixture-of-experts (MoE) learning for zero-shot inference or few-shot adaptation. Structured domain knowledge guides dependency-aware planning, while validation evidence supports feedback-driven revision of affected decisions. Experiments on the real-world DeepSense 6G dataset demonstrate improved vehicle detection and beam prediction over the considered simulation and fusion baselines. A separate orchestration benchmark evaluates task interpretation, dependency reasoning, and feedback-driven replanning across diverse deployment requests, showing improved plan correctness with structured domain knowledge and validation feedback.
展开完整摘要收起摘要↓
Multi-modal integrated sensing and communication (ISAC) enables environmental perception and reliable connectivity for intelligent wireless networks. Data-driven multi-modal ISAC models depend heavily on annotated real-world data to learn relationships across sensing and wireless observations, thereby constraining scalable deployment. Although synthetic data generation reduces the burden, adapting existing simulation pipelines to a target deployment requires consistent scene, sensing, wireless, and learning configurations, while mismatches among these coupled components impair sim-to-real transferability. To address the challenge, we propose an agentic artificial intelligence (AI) framework for sim-to-real multi-modal ISAC, named AIMS. Given a natural-language deployment request specifying the target task, deployment conditions, and real-data budget, AIMS derives a deployment-specific sim-to-real configuration and coordinates its execution to produce a deployment-specific task model. A two-agent architecture coordinates scene construction with task learning. A scene construction agent generates geographically grounded, synchronized sensing and wireless records from shared physical states, while a scene understanding agent configures task-relevant modalities and mixture-of-experts (MoE) learning for zero-shot inference or few-shot adaptation. Structured domain knowledge guides dependency-aware planning, while validation evidence supports feedback-driven revision of affected decisions. Experiments on the real-world DeepSense 6G dataset demonstrate improved vehicle detection and beam prediction over the considered simulation and fusion baselines. A separate orchestration benchmark evaluates task interpretation, dependency reasoning, and feedback-driven replanning across diverse deployment requests, showing improved plan correctness with structured domain knowledge and validation feedback.
作者Minki Kang, Ryo Hachiuma, Shaokun Zhang, Subhashree Radhakrishnan, Yonggan Fu, Jindong Jiang, Mingjie Liu, Ehsan Hosseini-Asl, Yi Dong, Yu-Chiang Frank Wang, Byung-Kwan Lee
Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When the same TMAX-9B model serves as the verifier, pairwise verification performs best among the evaluated verification mechanisms. Distilling responses from the stronger verifier into TMAX-9B further improves Pass@1, while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reaches higher success at lower estimated token cost than generating more trajectories alone. Mid-Harness also improves performance across additional models, benchmarks, and harnesses. These findings identify action scaling as a promising target for test-time compute scaling in terminal agents.
展开完整摘要收起摘要↓
Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When the same TMAX-9B model serves as the verifier, pairwise verification performs best among the evaluated verification mechanisms. Distilling responses from the stronger verifier into TMAX-9B further improves Pass@1, while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reaches higher success at lower estimated token cost than generating more trajectories alone. Mid-Harness also improves performance across additional models, benchmarks, and harnesses. These findings identify action scaling as a promising target for test-time compute scaling in terminal agents.
Background. A component replacement in a language-model agent changes an execution trajectory, potentially altering later observations, resource use, and recovery opportunities. Different evidence is needed to assess its task-level benefit and the contribution of local decision quality. Methods. This critical scoping review maps 348 studies and examines 90 comparison records: 88 from 40 included studies and two from supplementary studies. Eight purposively selected cases structure the synthesis around the replaced decision, executed conditions, measurement comparability, controls, and remaining explanations. Results. Of 222 studies reporting local decision metrics, 142 also report measured task endpoints and 49 report proxies. These counts identify studies that report both types of measurement, without establishing that the measurements come from matched comparisons. Outcome Monitors reports a package-level completion gain whose attribution to detector quality remains limited; First-chunk selection reports a local improvement assessed against an offline proxy endpoint; Evidence-Carrying Termination reports fewer premature unsupported terminations and completion non-inferiority, without establishing completion superiority. Cross-case analysis identifies three candidate mechanisms involving recovery and disruption, intervention timing, and downstream use. Attribution and deployment depend on the comparison controls, label definitions, and information available to the controller. Conclusions. The review distinguishes the task-level benefit of a component replacement from the contribution of local decision quality and derives eight claim-specific reporting items. Neither online execution nor simultaneous gains in local and task metrics alone establish that better local decisions explain the task-level gain.
展开完整摘要收起摘要↓
Background. A component replacement in a language-model agent changes an execution trajectory, potentially altering later observations, resource use, and recovery opportunities. Different evidence is needed to assess its task-level benefit and the contribution of local decision quality. Methods. This critical scoping review maps 348 studies and examines 90 comparison records: 88 from 40 included studies and two from supplementary studies. Eight purposively selected cases structure the synthesis around the replaced decision, executed conditions, measurement comparability, controls, and remaining explanations. Results. Of 222 studies reporting local decision metrics, 142 also report measured task endpoints and 49 report proxies. These counts identify studies that report both types of measurement, without establishing that the measurements come from matched comparisons. Outcome Monitors reports a package-level completion gain whose attribution to detector quality remains limited; First-chunk selection reports a local improvement assessed against an offline proxy endpoint; Evidence-Carrying Termination reports fewer premature unsupported terminations and completion non-inferiority, without establishing completion superiority. Cross-case analysis identifies three candidate mechanisms involving recovery and disruption, intervention timing, and downstream use. Attribution and deployment depend on the comparison controls, label definitions, and information available to the controller. Conclusions. The review distinguishes the task-level benefit of a component replacement from the contribution of local decision quality and derives eight claim-specific reporting items. Neither online execution nor simultaneous gains in local and task metrics alone establish that better local decisions explain the task-level gain.
Compact regulatory DNA can free up space in vector payloads, reduce synthesis and assay burden, and expose which sequence features drive predicted activity. Yet most model-based nucleic-acid designers optimize fixed-length sequences through substitutions; they do not ask which bases of an existing functional element can be removed while retaining predicted activity. We define the task of sequence slimming as selecting an exact-length, order-preserving subsequence while retaining activity. Modeled on the design benchmark NucleoBench, we propose a quantitative evaluation for slimming that balances sequence reduction with maintaining function. Each slimmer must return both the subsequence and its source indices, which can be used to verify that the slimmer obeyed task requirements. To our knowledge, this is the first dedicated benchmark of this deletion-only problem. The coding agent Empirical Research Assistant (ERA) then searched over executable designer programs. ERA received the task prompt and a successful substitution-only designer GrAdaBeam as a starting program, and it modified the designer to produce GRADASLIM. We report held-out evaluations for five transcription-factor binding targets, comparing random, greedy, and ERA-guided slimming at 400 and 100 bp. ERA has the highest mean in 9/10 settings. Paired bootstrap intervals for ERA minus greedy are above zero in all five 400-bp settings, below zero in one 100-bp setting, and overlap zero in the remaining four.
展开完整摘要收起摘要↓
Compact regulatory DNA can free up space in vector payloads, reduce synthesis and assay burden, and expose which sequence features drive predicted activity. Yet most model-based nucleic-acid designers optimize fixed-length sequences through substitutions; they do not ask which bases of an existing functional element can be removed while retaining predicted activity. We define the task of sequence slimming as selecting an exact-length, order-preserving subsequence while retaining activity. Modeled on the design benchmark NucleoBench, we propose a quantitative evaluation for slimming that balances sequence reduction with maintaining function. Each slimmer must return both the subsequence and its source indices, which can be used to verify that the slimmer obeyed task requirements. To our knowledge, this is the first dedicated benchmark of this deletion-only problem. The coding agent Empirical Research Assistant (ERA) then searched over executable designer programs. ERA received the task prompt and a successful substitution-only designer GrAdaBeam as a starting program, and it modified the designer to produce GRADASLIM. We report held-out evaluations for five transcription-factor binding targets, comparing random, greedy, and ERA-guided slimming at 400 and 100 bp. ERA has the highest mean in 9/10 settings. Paired bootstrap intervals for ERA minus greedy are above zero in all five 400-bp settings, below zero in one 100-bp setting, and overlap zero in the remaining four.
AI agents that control wallets read attacker-reachable content, so they can be steered into proposing harmful transactions. The usual last line of defense is a pre-signing check: a static allowlist, an LLM reviewer, or a transaction simulation. All three share a gap: the check describes the chain state at check time, but the transaction executes in a later state that an adversary can shape through front-running, contract upgrades or token-parameter changes. We call this state drift. We present Proof-Gated Signing (PGS), which simulates a proposed transaction, extracts its effects, and uses an SMT solver to check a declarative value-and-permission policy for every price in an oracle-uncertainty band. It then compiles on-chain post-conditions (wallet balance bounds, payee receipts, allowance caps and ownership) and proves that every execution satisfying them also satisfies the policy. The agent's smart-contract wallet enforces them atomically, so the guarantee applies to the executed transaction under arbitrary drift. On an open testbed of 260 scenarios (14 attack families including five drift and two adaptive families, and 12 benign families), with harm measured from attacker balances rather than from any policy, PGS prevented 93.6% of the 140 harmful scenarios and passed 97.5% of the benign ones. Simulation-only checking prevented 57.9% and a static allowlist 71.4%. None of the 50 drift scenarios produced attacker gain under PGS. The only unprevented family, an in-policy drain, was bounded by the per-session budget. We also find that giving an LLM reviewer a clean pre-drift simulation made it more likely to approve a drift attack. Overhead is about 41k gas and 0.1-0.2 s per check.
展开完整摘要收起摘要↓
AI agents that control wallets read attacker-reachable content, so they can be steered into proposing harmful transactions. The usual last line of defense is a pre-signing check: a static allowlist, an LLM reviewer, or a transaction simulation. All three share a gap: the check describes the chain state at check time, but the transaction executes in a later state that an adversary can shape through front-running, contract upgrades or token-parameter changes. We call this state drift. We present Proof-Gated Signing (PGS), which simulates a proposed transaction, extracts its effects, and uses an SMT solver to check a declarative value-and-permission policy for every price in an oracle-uncertainty band. It then compiles on-chain post-conditions (wallet balance bounds, payee receipts, allowance caps and ownership) and proves that every execution satisfying them also satisfies the policy. The agent's smart-contract wallet enforces them atomically, so the guarantee applies to the executed transaction under arbitrary drift. On an open testbed of 260 scenarios (14 attack families including five drift and two adaptive families, and 12 benign families), with harm measured from attacker balances rather than from any policy, PGS prevented 93.6% of the 140 harmful scenarios and passed 97.5% of the benign ones. Simulation-only checking prevented 57.9% and a static allowlist 71.4%. None of the 50 drift scenarios produced attacker gain under PGS. The only unprevented family, an in-policy drain, was bounded by the per-session budget. We also find that giving an LLM reviewer a clean pre-drift simulation made it more likely to approve a drift attack. Overhead is about 41k gas and 0.1-0.2 s per check.
Code generation has emerged as a central capability of large language models, with coding agents now able to produce functionally correct software projects from natural language prompts. However, functional correctness alone does not capture a critical dimension of generation quality: environment specification, defined as the accurate identification of the dependencies required to execute generated code, is equally critical. We develop an agent protocol for environment specification and introduce a three-layer framework comprising declared, runtime-installed, and necessary-and-sufficient dependencies to systematically assess coding agents for environment specification. Using this protocol, we evaluate the extent to which coding agents systematically misspecify software environment dependencies and how this misspecification varies across three agents, four languages, and fifty programming tasks. Our results show that current coding agents exhibit systematic generalization failures along this dimension, producing dependency specifications that are inconsistent, redundant, or incomplete in ways that functional tests do not detect. Across agents, dependency set agreement is as low as 7% for identical tasks, and newer agents show no meaningful improvement, suggesting the failure is not resolved by scale or recency. The largest divergence occurs between the declared and runtime dependency layers, implicating environment priors learned from the models' training distributions as the primary driver. Our findings establish environment specification as a distinct, measurable axis of code generation quality that current benchmarks do not capture, and motivate training objectives and evaluation protocols that jointly optimize for functional correctness and environmental portability.
展开完整摘要收起摘要↓
Code generation has emerged as a central capability of large language models, with coding agents now able to produce functionally correct software projects from natural language prompts. However, functional correctness alone does not capture a critical dimension of generation quality: environment specification, defined as the accurate identification of the dependencies required to execute generated code, is equally critical. We develop an agent protocol for environment specification and introduce a three-layer framework comprising declared, runtime-installed, and necessary-and-sufficient dependencies to systematically assess coding agents for environment specification. Using this protocol, we evaluate the extent to which coding agents systematically misspecify software environment dependencies and how this misspecification varies across three agents, four languages, and fifty programming tasks. Our results show that current coding agents exhibit systematic generalization failures along this dimension, producing dependency specifications that are inconsistent, redundant, or incomplete in ways that functional tests do not detect. Across agents, dependency set agreement is as low as 7% for identical tasks, and newer agents show no meaningful improvement, suggesting the failure is not resolved by scale or recency. The largest divergence occurs between the declared and runtime dependency layers, implicating environment priors learned from the models' training distributions as the primary driver. Our findings establish environment specification as a distinct, measurable axis of code generation quality that current benchmarks do not capture, and motivate training objectives and evaluation protocols that jointly optimize for functional correctness and environmental portability.
作者Jiayi Geng, Zhengxuan Wu, Kevin S. Chen, Seungone Kim, Joseph Janssen, Zora Zhiruo Wang, Bhupalee Kalita, Runtian Gao, Aaron Ho, Andrew Oakleigh Nelson, Olexandr Isayev, Francisco Villaescusa-Navarro, Ching-Yao Lai, Howard Chen, Graham Neubig
When Isaac Newton discovered the law of gravitation, he did so through an iterative process of analyzing observed data such as planetary patterns, finding the underlying mechanisms by describing patterns in mathematical equations, and refining his theory against the Moon's orbit, revealing the startling insight that the same force governs both falling apples and orbiting planets. Would it be possible for AI agents to make similar discoveries? To measure this ability, we introduce EurekaBench, a cross-domain benchmark that tests AI agents' ability to conduct long-horizon experiments and discover mechanisms that explain observations. We evaluate these mechanisms by the scientific insights that can be derived from them. EurekaBench contains an expert-verified set of 26 long-horizon tasks across neuroscience, computer science, chemistry, astrophysics, geophysics, and plasma physics, with a total of 306 scientific insights that the discovered mechanisms are expected to support. Our evaluation framework tests three axes of scientific discovery: agents' ability to follow known scientific constraints, the predictive accuracy of the discovered mechanisms, and whether these mechanisms yield scientific insights or inform future research. Our results show that current AI agents often overly fixate on predictive accuracy optimization, surpassing human scientists, while falling substantially short in deriving scientific insights.
展开完整摘要收起摘要↓
When Isaac Newton discovered the law of gravitation, he did so through an iterative process of analyzing observed data such as planetary patterns, finding the underlying mechanisms by describing patterns in mathematical equations, and refining his theory against the Moon's orbit, revealing the startling insight that the same force governs both falling apples and orbiting planets. Would it be possible for AI agents to make similar discoveries? To measure this ability, we introduce EurekaBench, a cross-domain benchmark that tests AI agents' ability to conduct long-horizon experiments and discover mechanisms that explain observations. We evaluate these mechanisms by the scientific insights that can be derived from them. EurekaBench contains an expert-verified set of 26 long-horizon tasks across neuroscience, computer science, chemistry, astrophysics, geophysics, and plasma physics, with a total of 306 scientific insights that the discovered mechanisms are expected to support. Our evaluation framework tests three axes of scientific discovery: agents' ability to follow known scientific constraints, the predictive accuracy of the discovered mechanisms, and whether these mechanisms yield scientific insights or inform future research. Our results show that current AI agents often overly fixate on predictive accuracy optimization, surpassing human scientists, while falling substantially short in deriving scientific insights.
Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously. Reward-wise normalization, as used in GDPO, preserves reward-specific relative information within rollout groups, but different objectives can still exhibit uneven learning progress. We study this behavior through advantage energy, the sum of a reward's squared advantages over a batch. Under idealized GDPO normalization, we show that this energy is proportional to active-group density: the fraction of rollout groups in which the reward provides nonzero relative advantages. This reveals a residual batch-level signal imbalance and provides a basis for calibrating reward contributions. Based on this relation, we propose Density-Aware Reward Aggregation (DARA). We derive an inverse-square-root density correction that gives greater weight to signals from less frequently active rewards. DARA computes its weights from each rollout batch, adapting to changes in reward activity throughout training without modifying the underlying policy optimization objective. Experiments on tool calling and mathematical reasoning show that DARA learns the targeted behaviors faster than GDPO, reaching high format compliance in up to 26% fewer training steps on tool calling and near-saturated length compliance in up to 65% fewer steps on mathematical reasoning, while remaining competitive in final performance. Our code is available at https://github.com/zhaihaotian/DARA.
展开完整摘要收起摘要↓
Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously. Reward-wise normalization, as used in GDPO, preserves reward-specific relative information within rollout groups, but different objectives can still exhibit uneven learning progress. We study this behavior through advantage energy, the sum of a reward's squared advantages over a batch. Under idealized GDPO normalization, we show that this energy is proportional to active-group density: the fraction of rollout groups in which the reward provides nonzero relative advantages. This reveals a residual batch-level signal imbalance and provides a basis for calibrating reward contributions. Based on this relation, we propose Density-Aware Reward Aggregation (DARA). We derive an inverse-square-root density correction that gives greater weight to signals from less frequently active rewards. DARA computes its weights from each rollout batch, adapting to changes in reward activity throughout training without modifying the underlying policy optimization objective. Experiments on tool calling and mathematical reasoning show that DARA learns the targeted behaviors faster than GDPO, reaching high format compliance in up to 26% fewer training steps on tool calling and near-saturated length compliance in up to 65% fewer steps on mathematical reasoning, while remaining competitive in final performance. Our code is available at https://github.com/zhaihaotian/DARA.
作者Andre Fu, Malik Drabla, Leon Liu, Meji Abidoye, Marek Suppa, Lata Mishra, Adnan El Assadi, Yiyuan Li
AI coding agents are ubiquitous in engineering workflows amongst industry and academia. Yet, despite their use in app coding, relatively less attention has been paid to their ability to execute on production incident response. This emerging field, termed agentic site-reliability-engineering (SRE) contains benchmarks limited by (1) unrealistic environments, typically toy repositories (2) non-standard framework implementations and (3) simple static verifiers. We introduce Incident-Arena, a human-built benchmark of 20 carefully selected tasks grounded in real-world deployed open source software. Each task deploys a production application to an ephemeral Kubernetes cluster, injecting a fault from the config layer through underlying images, and a sustained load profile given the task requirements. We also present a novel verification method, going beyond static checks to functional verifiers, holding systems level metrics stable, while ensuring repairs are done safely. Agent trials run an average of 2.81M tokens and 41 turns, going beyond existing benchmarks, demonstrating agentic long horizon reasoning. Across 20 tasks and 3 application substrates, frontier models score below 64.3%, with failures extending from diagnosis/localization errors, through incomplete repairs and unsafe regressions.
展开完整摘要收起摘要↓
AI coding agents are ubiquitous in engineering workflows amongst industry and academia. Yet, despite their use in app coding, relatively less attention has been paid to their ability to execute on production incident response. This emerging field, termed agentic site-reliability-engineering (SRE) contains benchmarks limited by (1) unrealistic environments, typically toy repositories (2) non-standard framework implementations and (3) simple static verifiers. We introduce Incident-Arena, a human-built benchmark of 20 carefully selected tasks grounded in real-world deployed open source software. Each task deploys a production application to an ephemeral Kubernetes cluster, injecting a fault from the config layer through underlying images, and a sustained load profile given the task requirements. We also present a novel verification method, going beyond static checks to functional verifiers, holding systems level metrics stable, while ensuring repairs are done safely. Agent trials run an average of 2.81M tokens and 41 turns, going beyond existing benchmarks, demonstrating agentic long horizon reasoning. Across 20 tasks and 3 application substrates, frontier models score below 64.3%, with failures extending from diagnosis/localization errors, through incomplete repairs and unsafe regressions.
作者Zhengyuan Jiang, Reachal Wang, Yuepeng Hu, Yupu Wang, Yuqi Jia, Neil Zhenqiang Gong
The performance of AI coding agents is highly dependent on their underlying coding rules. However, existing coding rules are typically hand-crafted and fixed, making the process labor-intensive and often suboptimal. In this work, we propose RuleEvolve, a self-evolving framework for coding rules. RuleEvolve maintains a pool of candidate coding rules and iteratively improves them. In each iteration, it employs an LLM-powered mutator module to generate variants from existing candidates, and then uses a judge module to evaluate these variants and update the pool with the best-performing ones. Extensive evaluations across two coding-agent frameworks, four backbone LLMs, and three benchmarks demonstrate that RuleEvolve outperforms both manual engineering and existing prompt optimization baselines in terms of functional correctness of the generated code, code length, and/or generation cost (e.g., tokens used).
展开完整摘要收起摘要↓
The performance of AI coding agents is highly dependent on their underlying coding rules. However, existing coding rules are typically hand-crafted and fixed, making the process labor-intensive and often suboptimal. In this work, we propose RuleEvolve, a self-evolving framework for coding rules. RuleEvolve maintains a pool of candidate coding rules and iteratively improves them. In each iteration, it employs an LLM-powered mutator module to generate variants from existing candidates, and then uses a judge module to evaluate these variants and update the pool with the best-performing ones. Extensive evaluations across two coding-agent frameworks, four backbone LLMs, and three benchmarks demonstrate that RuleEvolve outperforms both manual engineering and existing prompt optimization baselines in terms of functional correctness of the generated code, code length, and/or generation cost (e.g., tokens used).
Contextual security defenses prevent AI agents from taking rogue actions by synthesizing a task-specific policy and enforcing it on the agent's tool calls. In multi-step tasks, however, which actions are valid often depends on what the agent has already done and learned. We present Sapien, a policy engine for enforcing stateful contextual policies. A Sapien policy specifies permitted tool-call sequences using a regular expression extended with stateful predicates, deferred policy generation, and scoped semantic checks. We show that Sapien stays within a few percent of an unconstrained agent's utility. Even if the agent is fully hijacked, Sapien's policies rule out 93-95% of attacks on AgentDojo and 62-85% on Toolathlon (twice as many as tool allowlists on long-horizon tasks).
展开完整摘要收起摘要↓
Contextual security defenses prevent AI agents from taking rogue actions by synthesizing a task-specific policy and enforcing it on the agent's tool calls. In multi-step tasks, however, which actions are valid often depends on what the agent has already done and learned. We present Sapien, a policy engine for enforcing stateful contextual policies. A Sapien policy specifies permitted tool-call sequences using a regular expression extended with stateful predicates, deferred policy generation, and scoped semantic checks. We show that Sapien stays within a few percent of an unconstrained agent's utility. Even if the agent is fully hijacked, Sapien's policies rule out 93-95% of attacks on AgentDojo and 62-85% on Toolathlon (twice as many as tool allowlists on long-horizon tasks).
作者Rafi Ibn Sultan, Md. Sajid Alam Chowdhury, Saleh Zare Zade, Chengyin Li, Prashant Khanduri, Marco Brocanelli, Dongxiao Zhu
Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain. This early commitment constitutes premature discretization: an incorrect token selection can propagate errors through subsequent reasoning. We propose Soft Spatial Reasoning, a post-training framework that introduces soft thinking for spatial tasks in LVLMs. At each intermediate reasoning step, the LVLM forms a continuous soft state by mixing token embeddings rather than selecting a single token, allowing multiple candidate continuations to influence the next step. The appropriate degree of softness, however, can vary across reasoning steps: retaining multiple candidates may preserve a useful spatial interpretation, but if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. At the core of Soft Spatial Reasoning is AdaptSoft, a controller that uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step. To train AdaptSoft, we introduce a gradient-alignment learning objective that provides a step-specific learning signal for softness control without intermediate reasoning supervision. Across diverse spatial benchmarks, Soft Spatial Reasoning outperforms hard and fixed-soft CoT baselines using the same backbone, as well as a range of existing LVLMs. The source code is available at https://github.com/rafiibnsultan/Soft_Spatial_Reasoning
展开完整摘要收起摘要↓
Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain. This early commitment constitutes premature discretization: an incorrect token selection can propagate errors through subsequent reasoning. We propose Soft Spatial Reasoning, a post-training framework that introduces soft thinking for spatial tasks in LVLMs. At each intermediate reasoning step, the LVLM forms a continuous soft state by mixing token embeddings rather than selecting a single token, allowing multiple candidate continuations to influence the next step. The appropriate degree of softness, however, can vary across reasoning steps: retaining multiple candidates may preserve a useful spatial interpretation, but if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. At the core of Soft Spatial Reasoning is AdaptSoft, a controller that uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step. To train AdaptSoft, we introduce a gradient-alignment learning objective that provides a step-specific learning signal for softness control without intermediate reasoning supervision. Across diverse spatial benchmarks, Soft Spatial Reasoning outperforms hard and fixed-soft CoT baselines using the same backbone, as well as a range of existing LVLMs. The source code is available at https://github.com/rafiibnsultan/Soft_Spatial_Reasoning
作者Yuanhe Zhang, Xinyao Zhou, Haoran Gao, Yuyao Zhang, Zhenhong Zhou, Fanyu Meng, Li Sun, Sen Su
Uncontrolled repetition can prolong autoregressive generation in large language models (LLMs) and enable resource consumption attacks. Prior analyses of repetitive generation have identified strongly activated features in intermediate and late layers. However, how uncontrolled repetition activity emerges and develops before becoming prominent in these layers remains insufficiently understood. In this paper, we investigate this question primarily in large vision-language models (LVLMs), which support a richer set of uncontrolled repetitions through both visual and textual inputs. We propose Tokenwise Residual Comparison (TRC), a method that identifies and localizes anomalies associated with repetition from residual dynamics during generation. TRC compares attention and multilayer perceptron writes to the residual stream across generated tokens to identify patterns associated with repetition. It then selectively suppresses coordinates in the residual stream at the identified layer. Experiments show that TRC effectively mitigates uncontrolled repetition, reducing loop rates by 57% on average. Our analysis further shows that repetition semantics emerge in shallow layers and propagate through the residual stream, disrupting normal representations. TRC also generalizes to large language models (LLMs) and large reasoning models (LRMs), where it consistently captures analogous repetition dynamics and achieves effective mitigation. Our work broadens the study of repetitive generation from its prominent internal representations to earlier opportunities for intervention, providing insights for mitigating resource consumption attacks.
展开完整摘要收起摘要↓
Uncontrolled repetition can prolong autoregressive generation in large language models (LLMs) and enable resource consumption attacks. Prior analyses of repetitive generation have identified strongly activated features in intermediate and late layers. However, how uncontrolled repetition activity emerges and develops before becoming prominent in these layers remains insufficiently understood. In this paper, we investigate this question primarily in large vision-language models (LVLMs), which support a richer set of uncontrolled repetitions through both visual and textual inputs. We propose Tokenwise Residual Comparison (TRC), a method that identifies and localizes anomalies associated with repetition from residual dynamics during generation. TRC compares attention and multilayer perceptron writes to the residual stream across generated tokens to identify patterns associated with repetition. It then selectively suppresses coordinates in the residual stream at the identified layer. Experiments show that TRC effectively mitigates uncontrolled repetition, reducing loop rates by 57% on average. Our analysis further shows that repetition semantics emerge in shallow layers and propagate through the residual stream, disrupting normal representations. TRC also generalizes to large language models (LLMs) and large reasoning models (LRMs), where it consistently captures analogous repetition dynamics and achieves effective mitigation. Our work broadens the study of repetitive generation from its prominent internal representations to earlier opportunities for intervention, providing insights for mitigating resource consumption attacks.