作者Haonan Dong, Qiguan Feng, Kehan Jiang, Haoran Ye, Xin Zhang, Guojie Song
Autonomous agents have rapidly matured as task executors and seen widespread deployment via harnesses such as OpenClaw. Safety concerns have rightly drawn growing research attention, and beneath them lie the values silently steering agent behavior. Existing value benchmarks, however, remain confined to LLMs, leaving agent values largely uncharted. From intuitive, empirical, and theoretical vantage points, we show that an agent's values diverge from those of its underlying LLM, and the agentic modality further introduces dataset-, evaluation-, and system-level challenges absent from text-only protocols. We close this gap with Agent-ValueBench, the first benchmark dedicated to agent values. It features 394 executable environments across 16 domains, offering 4,335 value-conflict tasks that cover 28 value systems and 332 dimensions. Every instance is co-synthesized through our purpose-built end-to-end pipeline and curated per-instance by professional psychologists. Each task ships with two pole-aligned golden trajectories whose checkpoints anchor a trajectory-level rubric-based judge. Benchmarking 14 frontier proprietary and open-weights models across 4 mainstream harnesses, we uncover three concerted findings. Agent values first manifest as a Value Tide of cross-model homogeneity beneath interpretable counter-currents. This tide bends non-additively under harness pull, and yet more decisively under deliberate steering via embedded skills. Together these results signal that the agent-alignment lever is shifting from classical model alignment and prompt steering toward harness alignment and skill steering.
展开完整摘要收起摘要↓
Autonomous agents have rapidly matured as task executors and seen widespread deployment via harnesses such as OpenClaw. Safety concerns have rightly drawn growing research attention, and beneath them lie the values silently steering agent behavior. Existing value benchmarks, however, remain confined to LLMs, leaving agent values largely uncharted. From intuitive, empirical, and theoretical vantage points, we show that an agent's values diverge from those of its underlying LLM, and the agentic modality further introduces dataset-, evaluation-, and system-level challenges absent from text-only protocols. We close this gap with Agent-ValueBench, the first benchmark dedicated to agent values. It features 394 executable environments across 16 domains, offering 4,335 value-conflict tasks that cover 28 value systems and 332 dimensions. Every instance is co-synthesized through our purpose-built end-to-end pipeline and curated per-instance by professional psychologists. Each task ships with two pole-aligned golden trajectories whose checkpoints anchor a trajectory-level rubric-based judge. Benchmarking 14 frontier proprietary and open-weights models across 4 mainstream harnesses, we uncover three concerted findings. Agent values first manifest as a Value Tide of cross-model homogeneity beneath interpretable counter-currents. This tide bends non-additively under harness pull, and yet more decisively under deliberate steering via embedded skills. Together these results signal that the agent-alignment lever is shifting from classical model alignment and prompt steering toward harness alignment and skill steering.
Over the years, research in system identification has provided a rich set of methods for learning dynamical models, together with well-established theoretical guarantees. In practice, however, the choice of model class, training algorithm, and hyperparameter tuning is still largely left to empirical trial-and-error, requiring substantial expert time and domain experience. Motivated by recent advances in agentic artificial intelligence, we present ASIA, a framework that delegates this iterative search to a large language model acting as an autonomous coding agent. Building on existing agentic platforms, ASIA closes the loop between hypothesis, implementation, and evaluation without human intervention, requiring only a plain-English description of the identification problem. We conduct an empirical study of ASIA on two system identification benchmarks and analyse the agent's search behaviour, the architectures and training strategies it discovers, and the quality of the resulting models. We also discuss the potential of the approach and its current limitations, including implicit test leakage, reduced methodological transparency, and reproducibility concerns.
展开完整摘要收起摘要↓
Over the years, research in system identification has provided a rich set of methods for learning dynamical models, together with well-established theoretical guarantees. In practice, however, the choice of model class, training algorithm, and hyperparameter tuning is still largely left to empirical trial-and-error, requiring substantial expert time and domain experience. Motivated by recent advances in agentic artificial intelligence, we present ASIA, a framework that delegates this iterative search to a large language model acting as an autonomous coding agent. Building on existing agentic platforms, ASIA closes the loop between hypothesis, implementation, and evaluation without human intervention, requiring only a plain-English description of the identification problem. We conduct an empirical study of ASIA on two system identification benchmarks and analyse the agent's search behaviour, the architectures and training strategies it discovers, and the quality of the resulting models. We also discuss the potential of the approach and its current limitations, including implicit test leakage, reduced methodological transparency, and reproducibility concerns.
Multi-agent LLM systems introduce a security risk in which sensitive information accessed by one agent can propagate through shared context and reappear in downstream outputs, even without explicit adversarial intent. We formalise this phenomenon as propagation amplification, where leakage risk increases across agent boundaries as sensitive content is repeatedly exposed to downstream generators. Existing defences, including prompt-based safeguards, static pattern matching, and LLM-as-judge filtering, are not designed for this setting: they either operate after generation, rely primarily on surface-form patterns, or add substantial latency without modelling the generation process itself. To resolve these issues, we propose PRISM, a real-time defence that treats credential leakage as a sequential risk accumulation problem during generation. At each decoding step, PRISM combines 16 signals spanning lexical, structural, information-theoretic, behavioural, and contextual features into a calibrated risk score, enabling per-token intervention through green, yellow, and red risk zones. Our central observation is that credential reproduction is often preceded by a measurable shift in generation dynamics, characterised by entropy collapse and increasing logit concentration. When combined with text-structural cues such as identifier-pattern detection, these temporal signals provide an early warning of leakage before a secret is fully reconstructed. Across a 2,000-task adversarial benchmark covering 13 attack categories and three pressure levels in a heterogeneous four-agent pipeline, PRISM achieves F1 = 0.832 with precision = 1.000 and recall = 0.712, while producing no observed leakage on our benchmark (0.0% task-level leak rate) and preserving output utility of 0.893. It substantially outperforms the strongest baseline, Span Tagger, which achieves F1 = 0.719 with a 15.0% task-level leak rate.
展开完整摘要收起摘要↓
Multi-agent LLM systems introduce a security risk in which sensitive information accessed by one agent can propagate through shared context and reappear in downstream outputs, even without explicit adversarial intent. We formalise this phenomenon as propagation amplification, where leakage risk increases across agent boundaries as sensitive content is repeatedly exposed to downstream generators. Existing defences, including prompt-based safeguards, static pattern matching, and LLM-as-judge filtering, are not designed for this setting: they either operate after generation, rely primarily on surface-form patterns, or add substantial latency without modelling the generation process itself. To resolve these issues, we propose PRISM, a real-time defence that treats credential leakage as a sequential risk accumulation problem during generation. At each decoding step, PRISM combines 16 signals spanning lexical, structural, information-theoretic, behavioural, and contextual features into a calibrated risk score, enabling per-token intervention through green, yellow, and red risk zones. Our central observation is that credential reproduction is often preceded by a measurable shift in generation dynamics, characterised by entropy collapse and increasing logit concentration. When combined with text-structural cues such as identifier-pattern detection, these temporal signals provide an early warning of leakage before a secret is fully reconstructed. Across a 2,000-task adversarial benchmark covering 13 attack categories and three pressure levels in a heterogeneous four-agent pipeline, PRISM achieves F1 = 0.832 with precision = 1.000 and recall = 0.712, while producing no observed leakage on our benchmark (0.0% task-level leak rate) and preserving output utility of 0.893. It substantially outperforms the strongest baseline, Span Tagger, which achieves F1 = 0.719 with a 15.0% task-level leak rate.
Current LLM agents are proficient at calling isolated APIs but struggle with the "last mile" of commercial software automation. In real-world scenarios, tools are not independent; they are atomic, interdependent, and prone to environmental noise. We introduce $ComplexMCP$, a benchmark designed to evaluate agents in these rigorous conditions. Built on the Model Context Protocol (MCP), $ComplexMCP$ provides over 300 meticulously tested tools derived from 7 stateful sandboxes, ranging from office suites to financial systems. Unlike existing datasets, our benchmark utilizes a seed-driven architecture to simulate dynamic environment states and unpredictable API failures, ensuring a deterministic yet diverse evaluation. We evaluate various LLMs across full-context and RAG paradigms, revealing a stark performance gap: even top-tier models fail to exceed a 60% success rate, far trailing human performance 90%. Granular trajectory analysis identifies three fundamental bottlenecks: (1) $tool retrieval saturation$ as action spaces scale; (2) $over-confidence$, where agents skip essential environment verifications; and (3) $strategic defeatism$, a tendency to rationalize failure rather than pursuing recovery. These findings underscore the insufficiency of current agents for interdependent workflows, positioning $ComplexMCP$ as a critical testbed for the next generation of resilient autonomous systems.
展开完整摘要收起摘要↓
Current LLM agents are proficient at calling isolated APIs but struggle with the "last mile" of commercial software automation. In real-world scenarios, tools are not independent; they are atomic, interdependent, and prone to environmental noise. We introduce $ComplexMCP$, a benchmark designed to evaluate agents in these rigorous conditions. Built on the Model Context Protocol (MCP), $ComplexMCP$ provides over 300 meticulously tested tools derived from 7 stateful sandboxes, ranging from office suites to financial systems. Unlike existing datasets, our benchmark utilizes a seed-driven architecture to simulate dynamic environment states and unpredictable API failures, ensuring a deterministic yet diverse evaluation. We evaluate various LLMs across full-context and RAG paradigms, revealing a stark performance gap: even top-tier models fail to exceed a 60% success rate, far trailing human performance 90%. Granular trajectory analysis identifies three fundamental bottlenecks: (1) $tool retrieval saturation$ as action spaces scale; (2) $over-confidence$, where agents skip essential environment verifications; and (3) $strategic defeatism$, a tendency to rationalize failure rather than pursuing recovery. These findings underscore the insufficiency of current agents for interdependent workflows, positioning $ComplexMCP$ as a critical testbed for the next generation of resilient autonomous systems.
作者Zike Yuan, Yukun Cao, Han Zhang, Jianzhi Yan, Le Liu, Cai ke, Yue Yu, Hui Wang, Ming Liu, Bing Qin
Graph reasoning agents operating from natural-language inputs must solve a coupled problem: they must reconstruct a structured graph instance from text, decide whether existing computational assets are sufficient, interact with tools under a strict execution protocol, and satisfy an external verifier that checks structured correctness rather than textual plausibility. Existing approaches usually improve either the instruction side or the tool side in isolation, which leaves unclear what should be updated after failure. We propose EGL-SCA, a verifier-centric dual-space framework that models a graph reasoning agent using two collaborative components: an instruction-side policy space for reasoning strategies, and a tool-side program space for executable algorithmic tools. Our central mechanism is structural credit assignment, which maps trajectory evidence to conditional updates, precisely routing failures to either prompt optimization or tool synthesis and repair. To provide sufficient learning signals for dual-space adaptation, we introduce a training distribution stratified by task family, coupled with a Pareto-style retention strategy to balance success, generality, and parsimony. Experiments on four graph reasoning benchmarks show that EGL-SCA achieves a state-of-the-art 92.0% average success rate. By effectively co-evolving instructions and tools, our framework significantly outperforms both pure-prompting and fixed-toolbox baselines.
展开完整摘要收起摘要↓
Graph reasoning agents operating from natural-language inputs must solve a coupled problem: they must reconstruct a structured graph instance from text, decide whether existing computational assets are sufficient, interact with tools under a strict execution protocol, and satisfy an external verifier that checks structured correctness rather than textual plausibility. Existing approaches usually improve either the instruction side or the tool side in isolation, which leaves unclear what should be updated after failure. We propose EGL-SCA, a verifier-centric dual-space framework that models a graph reasoning agent using two collaborative components: an instruction-side policy space for reasoning strategies, and a tool-side program space for executable algorithmic tools. Our central mechanism is structural credit assignment, which maps trajectory evidence to conditional updates, precisely routing failures to either prompt optimization or tool synthesis and repair. To provide sufficient learning signals for dual-space adaptation, we introduce a training distribution stratified by task family, coupled with a Pareto-style retention strategy to balance success, generality, and parsimony. Experiments on four graph reasoning benchmarks show that EGL-SCA achieves a state-of-the-art 92.0% average success rate. By effectively co-evolving instructions and tools, our framework significantly outperforms both pure-prompting and fixed-toolbox baselines.
Visual anomaly detection (VAD) is crucial in many real-world fields, such as industrial inspection, medical imaging, infrastructure monitoring, and remote sensing. However, the specific anomaly definitions, data modalities, and annotation standards across different domains make it difficult to transfer single-domain trained VAD models. Vision-language models (VLMs), pre-trained on large-scale cross-domain data, can perform visual perception under task instructions, offering a promising solution for cross-domain VAD. However, single-inference VLM judgments are unreliable, since they rely more on prior knowledge than on normal-sample references or fine-grained feature evidence. We therefore present AnomalyClaw, a training-free VAD agent that turns anomaly judgment into a multi-round refutation process. In each round, the agent proposes candidate anomalies and refutes each against normal-sample references, drawing on a 13-tool library for visual verification, reference parsing, and frozen expert probing. On the CrossDomainVAD-12 benchmark (12 datasets), AnomalyClaw achieves consistent macro-AUROC improvements over single-step direct inference with +6.23 pp on GPT-5.5, +7.93 pp on Seed2.0-lite, and +3.52 pp on Qwen3.5-VL-27B. We further introduce an optional verbalized self-evolution extension. It builds an online rulebook from internal-branch disagreement without oracle labels. On Qwen3.5-VL-27B, it delivers a +2.09 pp mean gain, comparable to a K = 10 oracle-label supervised baseline (+1.99 pp). These results show that agentic refutation improve anomaly understanding and reasoning of VLMs, rather than merely aggregating tool outputs.
展开完整摘要收起摘要↓
Visual anomaly detection (VAD) is crucial in many real-world fields, such as industrial inspection, medical imaging, infrastructure monitoring, and remote sensing. However, the specific anomaly definitions, data modalities, and annotation standards across different domains make it difficult to transfer single-domain trained VAD models. Vision-language models (VLMs), pre-trained on large-scale cross-domain data, can perform visual perception under task instructions, offering a promising solution for cross-domain VAD. However, single-inference VLM judgments are unreliable, since they rely more on prior knowledge than on normal-sample references or fine-grained feature evidence. We therefore present AnomalyClaw, a training-free VAD agent that turns anomaly judgment into a multi-round refutation process. In each round, the agent proposes candidate anomalies and refutes each against normal-sample references, drawing on a 13-tool library for visual verification, reference parsing, and frozen expert probing. On the CrossDomainVAD-12 benchmark (12 datasets), AnomalyClaw achieves consistent macro-AUROC improvements over single-step direct inference with +6.23 pp on GPT-5.5, +7.93 pp on Seed2.0-lite, and +3.52 pp on Qwen3.5-VL-27B. We further introduce an optional verbalized self-evolution extension. It builds an online rulebook from internal-branch disagreement without oracle labels. On Qwen3.5-VL-27B, it delivers a +2.09 pp mean gain, comparable to a K = 10 oracle-label supervised baseline (+1.99 pp). These results show that agentic refutation improve anomaly understanding and reasoning of VLMs, rather than merely aggregating tool outputs.
Production LLM coding agents drift over long sessions: they forget user-specified constraints, slip into mistakes the user already flagged, and confabulate prior agreements. White-box approaches such as persona vectors require model weights and so cannot be applied to closed APIs (Claude, GPT-4) that most users actually interact with. We present Nautilus Compass, a black-box persona drift detector and agent memory layer for production coding agents. The method operates entirely at the prompt-text layer: cosine similarity between user prompts and behavioral anchor texts, aggregated by a weighted top-k mean using BGE-m3 embeddings. Compass is, to our knowledge, the only public agent memory layer (among Mem0, Letta, Cognee, Zep, MemOS, smrti verified May 2026) that does not call an LLM at index time to extract facts or build a graph; raw conversation text is embedded directly. The system ships as a Claude Code plugin, an MCP 2024-11-05 A2A server (Cursor, Cline, Hermes), a CLI, and a REST API on one daemon, with a Merkle-chained audit log for tamper-evident anchor updates. On a held-out test set built from real Claude Code session traces and labeled by an independent LLM judge, Compass reaches ROC AUC 0.83 for drift detection. The embedded retrieval pipeline scores 56.6% on LongMemEval-S v0.8 and 44.4% on EverMemBench-Dynamic (n=500), topping the four published EverMemBench Table 4 baselines. LongMemEval-S 56.6% is ~30 points below recent white-box leaders (90+%); we treat that as the architectural ceiling of the no-extraction design. End-to-end reproduction cost is $3.50 (~14x cheaper than GPT-4o-judged stacks). A paired cross-vendor behavior A/B accompanies these numbers as preliminary system-level evidence. Code, anchors, frozen test data, and audit-log tooling are MIT-licensed at github.com/chunxiaoxx/nautilus-compass.
展开完整摘要收起摘要↓
Production LLM coding agents drift over long sessions: they forget user-specified constraints, slip into mistakes the user already flagged, and confabulate prior agreements. White-box approaches such as persona vectors require model weights and so cannot be applied to closed APIs (Claude, GPT-4) that most users actually interact with. We present Nautilus Compass, a black-box persona drift detector and agent memory layer for production coding agents. The method operates entirely at the prompt-text layer: cosine similarity between user prompts and behavioral anchor texts, aggregated by a weighted top-k mean using BGE-m3 embeddings. Compass is, to our knowledge, the only public agent memory layer (among Mem0, Letta, Cognee, Zep, MemOS, smrti verified May 2026) that does not call an LLM at index time to extract facts or build a graph; raw conversation text is embedded directly. The system ships as a Claude Code plugin, an MCP 2024-11-05 A2A server (Cursor, Cline, Hermes), a CLI, and a REST API on one daemon, with a Merkle-chained audit log for tamper-evident anchor updates. On a held-out test set built from real Claude Code session traces and labeled by an independent LLM judge, Compass reaches ROC AUC 0.83 for drift detection. The embedded retrieval pipeline scores 56.6% on LongMemEval-S v0.8 and 44.4% on EverMemBench-Dynamic (n=500), topping the four published EverMemBench Table 4 baselines. LongMemEval-S 56.6% is ~30 points below recent white-box leaders (90+%); we treat that as the architectural ceiling of the no-extraction design. End-to-end reproduction cost is $3.50 (~14x cheaper than GPT-4o-judged stacks). A paired cross-vendor behavior A/B accompanies these numbers as preliminary system-level evidence. Code, anchors, frozen test data, and audit-log tooling are MIT-licensed at github.com/chunxiaoxx/nautilus-compass.
Large Language Model (LLM)-based agents (e.g., OpenClaw) increasingly rely on reusable skill libraries to solve artifact-rich tasks such as document-centric workflows and data-intensive analysis. As these libraries grow, a few works have attempted to study the Retrieval-Augmented Execution (RAE), which often first retrieves some external skills and other knowledge, then compiles the context using retrieved skills, and finally executes the task. Existing works mainly focus on optimizing skill retrieval and task execution, and they pay little attention to how to effectively organize the selected skill evidence in a form that is compact, grounded, and immediately usable for the downstream executors to complete tasks. To fill this gap, we propose SkillRAE, a two-stage RAE approach focusing on skill-based context compilation, which consists of the offline and online stages. Specifically, in the offline indexing stage, it builds a multi-level skill graph over skill communities, skills, and reusable subunits, for capturing their relationships. In the online retrieval stage, it first performs skill-ranked retrieval with selected-subunit evidence export in the graph, and then applies rescue-aware compact compilation to recover the key evidence. Together, these components compile a coarse-ranked skill set into a task-specific context that is compact, grounded, and immediately usable. Experiments on two public benchmarks show that SkillRAE achieves a significant improvement over baselines for RAE. For example, on SkillsBench, it achieves an improvement of 11.7% over the SOTA method. Ablation studies further show that our context compilation is crucial, instead of a mere prompt addition.
展开完整摘要收起摘要↓
Large Language Model (LLM)-based agents (e.g., OpenClaw) increasingly rely on reusable skill libraries to solve artifact-rich tasks such as document-centric workflows and data-intensive analysis. As these libraries grow, a few works have attempted to study the Retrieval-Augmented Execution (RAE), which often first retrieves some external skills and other knowledge, then compiles the context using retrieved skills, and finally executes the task. Existing works mainly focus on optimizing skill retrieval and task execution, and they pay little attention to how to effectively organize the selected skill evidence in a form that is compact, grounded, and immediately usable for the downstream executors to complete tasks. To fill this gap, we propose SkillRAE, a two-stage RAE approach focusing on skill-based context compilation, which consists of the offline and online stages. Specifically, in the offline indexing stage, it builds a multi-level skill graph over skill communities, skills, and reusable subunits, for capturing their relationships. In the online retrieval stage, it first performs skill-ranked retrieval with selected-subunit evidence export in the graph, and then applies rescue-aware compact compilation to recover the key evidence. Together, these components compile a coarse-ranked skill set into a task-specific context that is compact, grounded, and immediately usable. Experiments on two public benchmarks show that SkillRAE achieves a significant improvement over baselines for RAE. For example, on SkillsBench, it achieves an improvement of 11.7% over the SOTA method. Ablation studies further show that our context compilation is crucial, instead of a mere prompt addition.
Time series analysis underpins forecasting, monitoring, and decision making in domains such as finance and weather, where solving a task often requires both numerical accuracy and contextual reasoning. Recent progress has moved from specialized neural predictors to approaches built on LLMs and foundation models that can reason over time series inputs and use external tools. However, most such systems remain execution-centric: they focus on solving the current instance but learn little from exploratory execution. This is especially limiting in verifiable numeric settings, where multiple candidate executions and tool-use procedures may all be task-valid yet differ sharply in quantitative quality, and where early success can trigger tool-prior collapse that suppresses further exploration. To address this limitation, we present TimeClaw, an exploratory execution learning framework that turns exploratory execution into reusable hierarchical distilled experience through a four-stage loop: Explore, Compare, Distill, and Reinject. TimeClaw combines metric-supervised exploratory execution learning, task-aware tool dropout, and hierarchical distilled experience for inference-time reinjection, while keeping the base model frozen and avoiding online test-time adaptation. In an MTBench-aligned evaluation with 17 tasks that span finance and weather prediction and reasoning tasks, TimeClaw delivers consistent gains over the baselines. These results suggest that, for scientific systems, the bottleneck is not only execution-time capability, but how exploratory experience is compared, distilled, and reused.
展开完整摘要收起摘要↓
Time series analysis underpins forecasting, monitoring, and decision making in domains such as finance and weather, where solving a task often requires both numerical accuracy and contextual reasoning. Recent progress has moved from specialized neural predictors to approaches built on LLMs and foundation models that can reason over time series inputs and use external tools. However, most such systems remain execution-centric: they focus on solving the current instance but learn little from exploratory execution. This is especially limiting in verifiable numeric settings, where multiple candidate executions and tool-use procedures may all be task-valid yet differ sharply in quantitative quality, and where early success can trigger tool-prior collapse that suppresses further exploration. To address this limitation, we present TimeClaw, an exploratory execution learning framework that turns exploratory execution into reusable hierarchical distilled experience through a four-stage loop: Explore, Compare, Distill, and Reinject. TimeClaw combines metric-supervised exploratory execution learning, task-aware tool dropout, and hierarchical distilled experience for inference-time reinjection, while keeping the base model frozen and avoiding online test-time adaptation. In an MTBench-aligned evaluation with 17 tasks that span finance and weather prediction and reasoning tasks, TimeClaw delivers consistent gains over the baselines. These results suggest that, for scientific systems, the bottleneck is not only execution-time capability, but how exploratory experience is compared, distilled, and reused.
作者Jinha Chung, Byeongjun Shin, Jiin Kim, Minsoo Rhu
LLM-based agents deliver state-of-the-art performance across tasks but incur high end-to-end latency on edge devices. We introduce Agent-X, a software-only, accuracy-preserving framework that accelerates both the prefill and decode stages of on-device agent workloads. Agent-X's two key components rewrite prompts to leverage prefix caching tailored to agent-specific input-token patterns and enable LLM-free speculative decoding for fast token generation with minimal overhead. On representative agentic workloads, Agent-X achieves a 1.61x end-to-end speedup in real systems with no accuracy loss and can be seamlessly integrated into existing on-device AI agents. To the best of our knowledge, ours is the first to systematically characterize and eliminate latency bottlenecks in on-device agents.
展开完整摘要收起摘要↓
LLM-based agents deliver state-of-the-art performance across tasks but incur high end-to-end latency on edge devices. We introduce Agent-X, a software-only, accuracy-preserving framework that accelerates both the prefill and decode stages of on-device agent workloads. Agent-X's two key components rewrite prompts to leverage prefix caching tailored to agent-specific input-token patterns and enable LLM-free speculative decoding for fast token generation with minimal overhead. On representative agentic workloads, Agent-X achieves a 1.61x end-to-end speedup in real systems with no accuracy loss and can be seamlessly integrated into existing on-device AI agents. To the best of our knowledge, ours is the first to systematically characterize and eliminate latency bottlenecks in on-device agents.
Benchmarks for coding agents increasingly measure source-level software repair, and cybersecurity benchmarks increasingly measure broad capture-the-flag performance. Classical binary reverse engineering remains less precisely specified: given only an executable, can an agent recover validation logic and produce an input, serial, artifact, or key generator accepted by the program? We introduce CrackMeBench, a benchmark for evaluating language-model agents on educational CrackMe-style reverse-engineering tasks. CrackMeBench focuses on deterministic binary validation problems with executable oracles, symbol-poor binaries, explicit local tool access, and externally scored submissions rather than free-form explanations. The v0 benchmark combines eight public calibration CrackMes with twelve generated main-score tasks built from seeded C, Rust, and Go templates, and agents run through an equal shell interface in a no-network Linux Docker sandbox with standard reverse-engineering tools. In a three-model evaluation with a five-minute budget and three scored submissions per task, pass@3 on the generated split is 11/12 tasks (92%) for GPT-5.5, 7/12 (58%) for Claude Opus 4.7, and 5/12 (42%) for Kimi K2. The harder generated half separates the models more sharply, with pass@3 of 5/6, 2/6, and 1/6, respectively; on the eight-task public calibration split, pass@3 is 3/8, 2/8, and 1/8. CrackMeBench records pass@1 and pass@3, scored submissions, wall-clock time, command traces, tool categories, provider-reported token usage, estimated cost, and qualitative failure labels, providing a reproducible testbed for measuring progress from source-code reasoning toward autonomous binary analysis while restricting scope to educational, purpose-built programs.
展开完整摘要收起摘要↓
Benchmarks for coding agents increasingly measure source-level software repair, and cybersecurity benchmarks increasingly measure broad capture-the-flag performance. Classical binary reverse engineering remains less precisely specified: given only an executable, can an agent recover validation logic and produce an input, serial, artifact, or key generator accepted by the program? We introduce CrackMeBench, a benchmark for evaluating language-model agents on educational CrackMe-style reverse-engineering tasks. CrackMeBench focuses on deterministic binary validation problems with executable oracles, symbol-poor binaries, explicit local tool access, and externally scored submissions rather than free-form explanations. The v0 benchmark combines eight public calibration CrackMes with twelve generated main-score tasks built from seeded C, Rust, and Go templates, and agents run through an equal shell interface in a no-network Linux Docker sandbox with standard reverse-engineering tools. In a three-model evaluation with a five-minute budget and three scored submissions per task, pass@3 on the generated split is 11/12 tasks (92%) for GPT-5.5, 7/12 (58%) for Claude Opus 4.7, and 5/12 (42%) for Kimi K2. The harder generated half separates the models more sharply, with pass@3 of 5/6, 2/6, and 1/6, respectively; on the eight-task public calibration split, pass@3 is 3/8, 2/8, and 1/8. CrackMeBench records pass@1 and pass@3, scored submissions, wall-clock time, command traces, tool categories, provider-reported token usage, estimated cost, and qualitative failure labels, providing a reproducible testbed for measuring progress from source-code reasoning toward autonomous binary analysis while restricting scope to educational, purpose-built programs.
作者George Wu, Nan Jing, Qing Yi, Chuan Hao, Ming Yang, Feng Chang, Yuan Wei, Jian Yang, Ran Tao, Bryan Dai
Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks demonstrate that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, while hybrid reward training further improves scaling effectiveness and stability across iterations. Code and data are available at https://github.com/george-QF/TMAS-code.
展开完整摘要收起摘要↓
Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks demonstrate that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, while hybrid reward training further improves scaling effectiveness and stability across iterations. Code and data are available at https://github.com/george-QF/TMAS-code.
作者Seth Karten, Joel Zhang, Tersoo Upaa, Ruirong Feng, Wenzhe Li, Chengshuai Shi, Chi Jin, Kiran Vodrahalli
Coding harnesses such as Claude Code and OpenHands wrap foundation models with tools, memory, and planning, but no equivalent exists for embodied agents' long-horizon partial-observability decision-making. We first report our Gemini Plays Pokemon (GPP) experiments. With iterative human-in-the-loop harness refinement, GPP became the first AI system to complete Pokemon Blue, Yellow Legacy on hard mode, and Crystal without a lost battle. In the hardest stages, the agent itself began iterating on its strategy through long-context memory, surfacing emergent self-improvement signals alongside human-in-the-loop refinement. Continual Harness removes the human fully from this loop: a reset-free self-improving harness for embodied agents that formalizes and automates what we observed. Starting from only a minimal environment interface, the agent alternates between acting and refining its own prompt, sub-agents, skills, and memory, drawing on any past trajectory data. Prompt-optimization methods require episode resets; Continual Harness adapts online within a single run. On Pokemon Red and Emerald across frontier models, Continual Harness starting from scratch substantially reduces button-press cost relative to the minimalist baseline and recovers a majority of the gap to a hand-engineered expert harness, with capability-dependent gains, despite starting from the same raw interface with no curated knowledge, no hand-crafted tools, and no domain scaffolding. We then close the loop with the model itself: an online process-reward co-learning loop, in which an open-source agent's rollouts through the refining harness are relabeled by a frontier teacher and used to update the model, drives sustained in-game milestone progress on Pokemon Red without resetting the environment between training iterations.
展开完整摘要收起摘要↓
Coding harnesses such as Claude Code and OpenHands wrap foundation models with tools, memory, and planning, but no equivalent exists for embodied agents' long-horizon partial-observability decision-making. We first report our Gemini Plays Pokemon (GPP) experiments. With iterative human-in-the-loop harness refinement, GPP became the first AI system to complete Pokemon Blue, Yellow Legacy on hard mode, and Crystal without a lost battle. In the hardest stages, the agent itself began iterating on its strategy through long-context memory, surfacing emergent self-improvement signals alongside human-in-the-loop refinement. Continual Harness removes the human fully from this loop: a reset-free self-improving harness for embodied agents that formalizes and automates what we observed. Starting from only a minimal environment interface, the agent alternates between acting and refining its own prompt, sub-agents, skills, and memory, drawing on any past trajectory data. Prompt-optimization methods require episode resets; Continual Harness adapts online within a single run. On Pokemon Red and Emerald across frontier models, Continual Harness starting from scratch substantially reduces button-press cost relative to the minimalist baseline and recovers a majority of the gap to a hand-engineered expert harness, with capability-dependent gains, despite starting from the same raw interface with no curated knowledge, no hand-crafted tools, and no domain scaffolding. We then close the loop with the model itself: an online process-reward co-learning loop, in which an open-source agent's rollouts through the refining harness are relabeled by a frontier teacher and used to update the model, drives sustained in-game milestone progress on Pokemon Red without resetting the environment between training iterations.
Agent-based modeling (ABM) has long been used in economics to study human behavior, and large language model (LLM) agents now enable new forms of social and economic simulation. While prior work has discovered strategic deception by LLM agents in financial trading and auction markets, e-commerce remains underexplored despite its distinctive information asymmetry: sellers privately observe product quality, whereas buyers rely on advertised claims and reputation signals. We introduce TruthMarketTwin, a controlled simulation framework for studying LLM-agent behavior in e-commerce markets. The framework is one of the first to model bilateral trade under asymmetric information sharing, where agents make strategic listing, purchasing, rating, and recourse-related decisions to optimize seller profit and buyer utility. We find that LLM agents released into traditional markets autonomously exploit weaknesses in reputation-based governance, while warrant enforcement reduces deception and reshapes strategic reasoning. Our results position LLM-agent simulation as a tool for studying institution-governed autonomous markets.
展开完整摘要收起摘要↓
Agent-based modeling (ABM) has long been used in economics to study human behavior, and large language model (LLM) agents now enable new forms of social and economic simulation. While prior work has discovered strategic deception by LLM agents in financial trading and auction markets, e-commerce remains underexplored despite its distinctive information asymmetry: sellers privately observe product quality, whereas buyers rely on advertised claims and reputation signals. We introduce TruthMarketTwin, a controlled simulation framework for studying LLM-agent behavior in e-commerce markets. The framework is one of the first to model bilateral trade under asymmetric information sharing, where agents make strategic listing, purchasing, rating, and recourse-related decisions to optimize seller profit and buyer utility. We find that LLM agents released into traditional markets autonomously exploit weaknesses in reputation-based governance, while warrant enforcement reduces deception and reshapes strategic reasoning. Our results position LLM-agent simulation as a tool for studying institution-governed autonomous markets.
作者Haoyu Huang, Jiaxin Bai, Shujie Liu, Yang Wei, Hong Ting Tsang, Yisen Gao, Zhongwei Xie, Yufei Li, Yangqiu Song
Agent-compiled knowledge bases provide persistent external knowledge for large language model (LLM) agents in open-ended, knowledge-intensive downstream tasks. Yet their quality is systematically limited by incompleteness, incorrectness, and redundancy, manifested as missing evidence or cross-document links, low-confidence or imprecise claims, and ambiguous or coreference resolution issues. Such defects compound under iterative use, degrading retrieval fidelity and downstream task performance. We present DeepRefine, a general LLM-based reasoning model for agent-compiled knowledge refinement that improves the quality of any pre-constructed knowledge bases with user queries to make it more suitable for the downstream tasks. DeepRefine performs multi-turn interactions with the knowledge base and conducts abductive diagnosis over interaction history, localizes likely defects, and executes targeted refinement actions for incremental knowledge base updates. To optimize refinement policies of DeepRefine without gold references, we introduce a Gain-Beyond-Draft (GBD) reward and train the reasoning process end-to-end via reinforcement learning. Extensive experiments demonstrate consistent downstream gains over strong baselines.
展开完整摘要收起摘要↓
Agent-compiled knowledge bases provide persistent external knowledge for large language model (LLM) agents in open-ended, knowledge-intensive downstream tasks. Yet their quality is systematically limited by incompleteness, incorrectness, and redundancy, manifested as missing evidence or cross-document links, low-confidence or imprecise claims, and ambiguous or coreference resolution issues. Such defects compound under iterative use, degrading retrieval fidelity and downstream task performance. We present DeepRefine, a general LLM-based reasoning model for agent-compiled knowledge refinement that improves the quality of any pre-constructed knowledge bases with user queries to make it more suitable for the downstream tasks. DeepRefine performs multi-turn interactions with the knowledge base and conducts abductive diagnosis over interaction history, localizes likely defects, and executes targeted refinement actions for incremental knowledge base updates. To optimize refinement policies of DeepRefine without gold references, we introduce a Gain-Beyond-Draft (GBD) reward and train the reasoning process end-to-end via reinforcement learning. Extensive experiments demonstrate consistent downstream gains over strong baselines.
Self-evolving language-model agents must decide what to learn next and how to preserve what they have learned across iterations. Existing systems typically carry this cross-iteration knowledge as natural-language feedback, flat episodic memory, or implicit reinforcement signals, none of which cleanly supports a frozen weak backbone at inference time. This paper introduces MAGE (Multi-Agent Graph-guided Evolution), a framework that externalizes self-knowledge into a four-subgraph co-evolutionary knowledge graph. Its experience subgraph stores both teacher-written failure corrections and the learner's own past correct reasoning traces, which are retrieved as task-conditioned guidance for a frozen execution model. During evolution, the graph, a task-level search bandit, and a skill-level routing bandit are updated from the same reward stream, while the learner's backbone remains unchanged. We further provide structural analysis showing how append-only memory growth, bounded curriculum coverage, and task-filtered retrieval together support stable improvement of the retrieval substrate for frozen-learner evolution. Across nine benchmarks spanning mathematical reasoning, multi-hop and open-domain question answering, spatio-temporal analysis, financial numerical reasoning, medical multiple-choice, an open-world survival game, and web navigation, MAGE achieves strong performance against prompt-based frozen-backbone baselines. Ablations show that self-harvested success traces and teacher-written corrections are complementary, with success memories contributing most on reasoning-template-heavy tasks and corrective memories supporting harder composition and interaction settings.
展开完整摘要收起摘要↓
Self-evolving language-model agents must decide what to learn next and how to preserve what they have learned across iterations. Existing systems typically carry this cross-iteration knowledge as natural-language feedback, flat episodic memory, or implicit reinforcement signals, none of which cleanly supports a frozen weak backbone at inference time. This paper introduces MAGE (Multi-Agent Graph-guided Evolution), a framework that externalizes self-knowledge into a four-subgraph co-evolutionary knowledge graph. Its experience subgraph stores both teacher-written failure corrections and the learner's own past correct reasoning traces, which are retrieved as task-conditioned guidance for a frozen execution model. During evolution, the graph, a task-level search bandit, and a skill-level routing bandit are updated from the same reward stream, while the learner's backbone remains unchanged. We further provide structural analysis showing how append-only memory growth, bounded curriculum coverage, and task-filtered retrieval together support stable improvement of the retrieval substrate for frozen-learner evolution. Across nine benchmarks spanning mathematical reasoning, multi-hop and open-domain question answering, spatio-temporal analysis, financial numerical reasoning, medical multiple-choice, an open-world survival game, and web navigation, MAGE achieves strong performance against prompt-based frozen-backbone baselines. Ablations show that self-harvested success traces and teacher-written corrections are complementary, with success memories contributing most on reasoning-template-heavy tasks and corrective memories supporting harder composition and interaction settings.
作者Deepanshu, Divi Saxena, Deepali Rana, Ayesha Varshney, Sahinur Rahman Laskar
Legal information in India remains largely inaccessible due to the complexity of legal language and the sheer volume of legal documentation involved in research and case analysis. This paper presents NyayaAI, an AI-powered legal assistant that automates and simplifies legal workflows for lawyers, law students, and general users. The system combines Large Language Models with a Retrieval-Augmented Generation pipeline grounded in a curated Indian legal knowledge base comprising constitutional provisions, statutes, case laws, and judicial precedents. A multi-agent architecture orchestrated through the Mastra TypeScript framework coordinates a main agent with specialized sub-agents handling legal research, document summarization, case law retrieval, and drafting assistance. A compliance module validates all responses before delivery. Domain classification achieved 70% precision across test samples, with RAG retrieval precision at 74% and overall response accuracy at 72%, demonstrating that structured multi-agent LLM systems can meaningfully improve legal accessibility and workflow efficiency. The code\footnote{https://github.com/B97784/NyayaAI} is made publicly available for the benefit of the research community.
展开完整摘要收起摘要↓
Legal information in India remains largely inaccessible due to the complexity of legal language and the sheer volume of legal documentation involved in research and case analysis. This paper presents NyayaAI, an AI-powered legal assistant that automates and simplifies legal workflows for lawyers, law students, and general users. The system combines Large Language Models with a Retrieval-Augmented Generation pipeline grounded in a curated Indian legal knowledge base comprising constitutional provisions, statutes, case laws, and judicial precedents. A multi-agent architecture orchestrated through the Mastra TypeScript framework coordinates a main agent with specialized sub-agents handling legal research, document summarization, case law retrieval, and drafting assistance. A compliance module validates all responses before delivery. Domain classification achieved 70% precision across test samples, with RAG retrieval precision at 74% and overall response accuracy at 72%, demonstrating that structured multi-agent LLM systems can meaningfully improve legal accessibility and workflow efficiency. The code\footnote{https://github.com/B97784/NyayaAI} is made publicly available for the benefit of the research community.
Large language and vision-language models increasingly power agents that act on a user's behalf through command-line interface (CLI) harnesses. However, most agent benchmarks still rely on synthetic sandboxes, short-horizon tasks, mock-service APIs, and final-answer checks, leaving open whether agents can complete realistic long-horizon work in the runtimes where they are deployed. This work presents WildClawBench, a native-runtime benchmark of 60 human-authored, bilingual, multimodal tasks spanning six thematic categories. Each task averages roughly 8 minutes of wall-clock time and over 20 tool calls, and runs inside a reproducible Docker container hosting an actual CLI agent harness (OpenClaw, Claude Code, Codex, or Hermes Agent) with access to real tools rather than mock services. Grading is hybrid, combining deterministic rule-based checks, environment-state auditing of side effects, and an LLM/VLM judge for semantic verification. Across 19 frontier models, the best, Claude Opus 4.7, reaches only 62.2% overall under OpenClaw, while every other model stays below 60%, and switching harness alone shifts a single model by up to 18 points. These results show that long-horizon, native-runtime agent evaluation remains a far-from-resolved task for current frontier models. We release the tasks, code, and containerized tooling to support reproducible evaluation.
展开完整摘要收起摘要↓
Large language and vision-language models increasingly power agents that act on a user's behalf through command-line interface (CLI) harnesses. However, most agent benchmarks still rely on synthetic sandboxes, short-horizon tasks, mock-service APIs, and final-answer checks, leaving open whether agents can complete realistic long-horizon work in the runtimes where they are deployed. This work presents WildClawBench, a native-runtime benchmark of 60 human-authored, bilingual, multimodal tasks spanning six thematic categories. Each task averages roughly 8 minutes of wall-clock time and over 20 tool calls, and runs inside a reproducible Docker container hosting an actual CLI agent harness (OpenClaw, Claude Code, Codex, or Hermes Agent) with access to real tools rather than mock services. Grading is hybrid, combining deterministic rule-based checks, environment-state auditing of side effects, and an LLM/VLM judge for semantic verification. Across 19 frontier models, the best, Claude Opus 4.7, reaches only 62.2% overall under OpenClaw, while every other model stays below 60%, and switching harness alone shifts a single model by up to 18 points. These results show that long-horizon, native-runtime agent evaluation remains a far-from-resolved task for current frontier models. We release the tasks, code, and containerized tooling to support reproducible evaluation.
作者Shijue Huang, Hangyu Guo, Chenxin Li, Junting Lu, Xinyu Geng, Zhaochen Su, Zhenyu Li, Shuang Chen, Hongru Wang, Yi R. Fung
Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
展开完整摘要收起摘要↓
Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
Multimodal large language models increasingly solve vision-centric tasks by calling external tools for visual inspection, OCR, retrieval, calculation, and multi-step reasoning. Current tool-using agents usually expose the executed tool trajectory and the final answer, but they rarely specify which tool observation supports each generated claim. We call this missing claim-level dependency structure the provenance gap. The gap makes tool use hard to verify and hard to optimize, because useful evidence, redundant exploration, and unsupported reasoning are mixed in the same trajectory. We introduce TRACER, a framework for verifiable generative provenance in multimodal tool-using agents. Instead of adding citations after generation, TRACER generates each answer sentence together with a structured provenance record that identifies the supporting tool turn, evidence unit, and semantic support relation. Its relation space contains Quotation, Compression, and Inference, covering direct reuse, faithful condensation, and grounded derivation. TRACER verifies each record through schema checking, tool-turn alignment, source authenticity, and relation rationality, and then converts verified provenance into traceability constraints and provenance-derived local credit for reinforcement learning. We further construct TRACE-Bench, a benchmark for sentence-level provenance reconstruction from coarse multimodal tool trajectories. On TRACE-Bench, simply adding tools often introduces noise. With Qwen3-VL-8B, TRACER reaches 78.23% answer accuracy and 95.72% summary accuracy, outperforming the strongest closed-source tool-augmented baseline by 23.80 percentage points. Compared with tool-only supervised fine-tuning, it also reduces total test-set tool calls from 4949 to 3486. These results show that reliable multimodal tool reasoning depends on provenance-aware use of observations, not on more tool calls alone.
展开完整摘要收起摘要↓
Multimodal large language models increasingly solve vision-centric tasks by calling external tools for visual inspection, OCR, retrieval, calculation, and multi-step reasoning. Current tool-using agents usually expose the executed tool trajectory and the final answer, but they rarely specify which tool observation supports each generated claim. We call this missing claim-level dependency structure the provenance gap. The gap makes tool use hard to verify and hard to optimize, because useful evidence, redundant exploration, and unsupported reasoning are mixed in the same trajectory. We introduce TRACER, a framework for verifiable generative provenance in multimodal tool-using agents. Instead of adding citations after generation, TRACER generates each answer sentence together with a structured provenance record that identifies the supporting tool turn, evidence unit, and semantic support relation. Its relation space contains Quotation, Compression, and Inference, covering direct reuse, faithful condensation, and grounded derivation. TRACER verifies each record through schema checking, tool-turn alignment, source authenticity, and relation rationality, and then converts verified provenance into traceability constraints and provenance-derived local credit for reinforcement learning. We further construct TRACE-Bench, a benchmark for sentence-level provenance reconstruction from coarse multimodal tool trajectories. On TRACE-Bench, simply adding tools often introduces noise. With Qwen3-VL-8B, TRACER reaches 78.23% answer accuracy and 95.72% summary accuracy, outperforming the strongest closed-source tool-augmented baseline by 23.80 percentage points. Compared with tool-only supervised fine-tuning, it also reduces total test-set tool calls from 4949 to 3486. These results show that reliable multimodal tool reasoning depends on provenance-aware use of observations, not on more tool calls alone.
作者Xinrun Wang, Chang Yang, He Zhao, Zhuoyi Lin, Shuyue Hu
LLM-based foundation agents that perceive, reason, and act across thousands of reasoning steps are rapidly becoming the dominant paradigm for deploying artificial intelligence in open-ended, long-horizon complex tasks. Despite this significance, the field remains overwhelmingly engineering-driven. Engineering practice has converged on useful primitives (tool loops, memory banks, harnesses, reflection steps), yet these are assembled by empirical trial and error rather than from first principles. Fundamental questions remain open: under what conditions does a long-running agent remain on-task? How should an agent respond when its environment exceeds its representational capacity? What architectural properties are necessary for safe self-improvement? We argue that cybernetics, the mid-twentieth-century science of control and communication in complex systems, provides the missing theoretical scaffold for foundation agents. By mapping six canonical laws of classical cybernetics onto six agent design principles, and synthesizing those principles into three engineering desiderata (reliability, lifelong running, and self-Improvement), we arrive at a framework termed Agent Cybernetics. Three application domains, code generation, computer use and automated research, exemplify the analytical framework of agent cybernetics by identifying failure modes and concrete engineering recommendations. We hope that agent cybernetics opens a new research venue and establishes the scientific foundation that foundation agents need for principled, reliable real-world deployment.
展开完整摘要收起摘要↓
LLM-based foundation agents that perceive, reason, and act across thousands of reasoning steps are rapidly becoming the dominant paradigm for deploying artificial intelligence in open-ended, long-horizon complex tasks. Despite this significance, the field remains overwhelmingly engineering-driven. Engineering practice has converged on useful primitives (tool loops, memory banks, harnesses, reflection steps), yet these are assembled by empirical trial and error rather than from first principles. Fundamental questions remain open: under what conditions does a long-running agent remain on-task? How should an agent respond when its environment exceeds its representational capacity? What architectural properties are necessary for safe self-improvement? We argue that cybernetics, the mid-twentieth-century science of control and communication in complex systems, provides the missing theoretical scaffold for foundation agents. By mapping six canonical laws of classical cybernetics onto six agent design principles, and synthesizing those principles into three engineering desiderata (reliability, lifelong running, and self-Improvement), we arrive at a framework termed Agent Cybernetics. Three application domains, code generation, computer use and automated research, exemplify the analytical framework of agent cybernetics by identifying failure modes and concrete engineering recommendations. We hope that agent cybernetics opens a new research venue and establishes the scientific foundation that foundation agents need for principled, reliable real-world deployment.
作者Raphael Trumpp, Ömer Veysel Çağatan, Barış Akgün, Marco Caccamo
Pixel-based deep reinforcement learning agents are typically trained on heavily downsampled visual observations, a convention inherited from early benchmarks rather than grounded in principled design. In this work, we show that observation resolution is a critical yet overlooked variable for policy learning: higher-resolution inputs can substantially improve both performance and generalization, provided the network architecture can process them effectively. We find that the widely used Impala encoder, which flattens spatial features into a vector, suffers from quadratic parameter growth as resolution increases and fails to leverage the additional visual detail. Replacing this operation with global average pooling, as in the Impoola architecture, decouples parameter count from resolution and yields consistent improvements across resolutions and network widths - at their respective best conditions, visual scaling unlocks a 28 % performance gain for Impoola over Impala. These gains are strongest in environments that require precise perception of small or distant objects, and gradient saliency analysis confirms that the underlying mechanism is a more spatially localized visual attention of the policy at higher resolutions. Our results challenge the prevailing practice of aggressive input downsampling and position resolution-independent architectures as a simple, effective path toward scalable visual deep RL. To facilitate future research on resolution scaling in deep RL, we publicly release the open-source code for the Procgen-HD benchmark: https://github.com/raphajaner/procgen-hd.
展开完整摘要收起摘要↓
Pixel-based deep reinforcement learning agents are typically trained on heavily downsampled visual observations, a convention inherited from early benchmarks rather than grounded in principled design. In this work, we show that observation resolution is a critical yet overlooked variable for policy learning: higher-resolution inputs can substantially improve both performance and generalization, provided the network architecture can process them effectively. We find that the widely used Impala encoder, which flattens spatial features into a vector, suffers from quadratic parameter growth as resolution increases and fails to leverage the additional visual detail. Replacing this operation with global average pooling, as in the Impoola architecture, decouples parameter count from resolution and yields consistent improvements across resolutions and network widths - at their respective best conditions, visual scaling unlocks a 28 % performance gain for Impoola over Impala. These gains are strongest in environments that require precise perception of small or distant objects, and gradient saliency analysis confirms that the underlying mechanism is a more spatially localized visual attention of the policy at higher resolutions. Our results challenge the prevailing practice of aggressive input downsampling and position resolution-independent architectures as a simple, effective path toward scalable visual deep RL. To facilitate future research on resolution scaling in deep RL, we publicly release the open-source code for the Procgen-HD benchmark: https://github.com/raphajaner/procgen-hd.
作者Zichao Yu, Shengze Xu, Bingqing Jiang, Wenyi Zhang, Difan Zou
Diffusion large language models (dLLMs) offer a promising route to parallel and efficient text generation, but improving their reasoning ability requires effective post-training. Reinforcement learning with verifiable rewards (RLVR) is a natural choice for this purpose, yet its application to dLLMs is hindered by the absence of tractable sequence-level log-ratios, which are central to standard policy optimization. The lack of tractable sequence-level log-ratios forces existing methods to rely on high-variance ELBO-based approximations, where high verifier rewards can amplify inaccurate score estimates and destabilize RL training. To overcome this issue, we propose Relative Score Policy Optimization (RSPO), a simple RLVR method that uses verifiable rewards to calibrate noisy likelihood estimates in dLLMs. The core of our algorithm relies on a key observation: a reward advantage can be interpreted not only as an update direction, but also as a target for the relative log-ratio between the current and reference policies. Accordingly, RSPO calibrates this noisy relative log-ratio estimate by comparing its reward advantage with the reward-implied target relative log-ratio, updating the policy according to the gap between the current estimate and the target rather than the raw advantage alone. Experiments on mathematical reasoning and planning benchmarks show that RSPO yields especially strong gains on planning tasks and competitive mathematical-reasoning performance.
展开完整摘要收起摘要↓
Diffusion large language models (dLLMs) offer a promising route to parallel and efficient text generation, but improving their reasoning ability requires effective post-training. Reinforcement learning with verifiable rewards (RLVR) is a natural choice for this purpose, yet its application to dLLMs is hindered by the absence of tractable sequence-level log-ratios, which are central to standard policy optimization. The lack of tractable sequence-level log-ratios forces existing methods to rely on high-variance ELBO-based approximations, where high verifier rewards can amplify inaccurate score estimates and destabilize RL training. To overcome this issue, we propose Relative Score Policy Optimization (RSPO), a simple RLVR method that uses verifiable rewards to calibrate noisy likelihood estimates in dLLMs. The core of our algorithm relies on a key observation: a reward advantage can be interpreted not only as an update direction, but also as a target for the relative log-ratio between the current and reference policies. Accordingly, RSPO calibrates this noisy relative log-ratio estimate by comparing its reward advantage with the reward-implied target relative log-ratio, updating the policy according to the gap between the current estimate and the target rather than the raw advantage alone. Experiments on mathematical reasoning and planning benchmarks show that RSPO yields especially strong gains on planning tasks and competitive mathematical-reasoning performance.
作者Zijun Shen, Sihan Yang, Ruichuan An, Ziyu Guo, Hao Liang, Ming Lu, Renrui Zhang, Wentao Zhang
Unified Multimodal Models (UMMs) excel in general tasks but struggle to bridge the gap between personalized understanding and generation. Prior works largely rely on implicit token-level alignment via supervised fine-tuning, which fails to fully capture the potential synergy between comprehension and creation. In this work, we propose Sync-R1, an end-to-end reinforcement learning framework that jointly optimizes personalized understanding and generation within a single, explicit reasoning loop. Through this unified feedback process, Sync-R1 enables personalized comprehension to guide content creation, while the resulting generation quality reciprocally refines understanding within an integrated reward landscape. To efficiently orchestrate this dual-task synergy, we introduce Sync-GRPO, a reinforcement learning method utilizing an ensemble reward system. Furthermore, we propose Dynamic Group Scaling (DGS), which adaptively filters low-potential trajectories to reduce gradient variance and accelerate convergence. To better reflect real-world complexity, we introduce UnifyBench++, featuring denser textual descriptions and richer user contexts. Experimental results demonstrate that Sync-R1 achieves state-of-the-art performance, showcasing superior cross-task reasoning and robust personalization without requiring complex cold-start procedures. The code and the UnifyBench++ dataset will be released at: https://github.com/arctanxarc/UniCTokens.
展开完整摘要收起摘要↓
Unified Multimodal Models (UMMs) excel in general tasks but struggle to bridge the gap between personalized understanding and generation. Prior works largely rely on implicit token-level alignment via supervised fine-tuning, which fails to fully capture the potential synergy between comprehension and creation. In this work, we propose Sync-R1, an end-to-end reinforcement learning framework that jointly optimizes personalized understanding and generation within a single, explicit reasoning loop. Through this unified feedback process, Sync-R1 enables personalized comprehension to guide content creation, while the resulting generation quality reciprocally refines understanding within an integrated reward landscape. To efficiently orchestrate this dual-task synergy, we introduce Sync-GRPO, a reinforcement learning method utilizing an ensemble reward system. Furthermore, we propose Dynamic Group Scaling (DGS), which adaptively filters low-potential trajectories to reduce gradient variance and accelerate convergence. To better reflect real-world complexity, we introduce UnifyBench++, featuring denser textual descriptions and richer user contexts. Experimental results demonstrate that Sync-R1 achieves state-of-the-art performance, showcasing superior cross-task reasoning and robust personalization without requiring complex cold-start procedures. The code and the UnifyBench++ dataset will be released at: https://github.com/arctanxarc/UniCTokens.
Recent advancements in Neural Machine Translation (NMT) have significantly improved translation quality. However, the increasing size and complexity of state-of-the-art models present significant challenges for deployment on resource-limited devices. Knowledge distillation (KD) is a promising approach for compressing models, but its effectiveness diminishes when there is a large capacity gap between teacher and student models. To address this issue, we propose Evolving Knowledge Distillation (EKD), a progressive training framework in which the student model learns from a sequence of teachers with gradually increasing capacities. Experiments on IWSLT-14, WMT-17, and WMT-23 benchmarks show that EKD leads to consistent improvements at each stage. On IWSLT-14, the final student achieves a BLEU score of 34.24, narrowing the gap to the strongest teacher (34.32 BLEU) to just 0.08 BLEU. Similar trends are observed on other datasets. These results demonstrate that EKD effectively bridges the capacity gap, enabling compact models to achieve performance close to that of much larger teacher models.Code and models are available at https://github.com/agi-content-generation/EKD.
展开完整摘要收起摘要↓
Recent advancements in Neural Machine Translation (NMT) have significantly improved translation quality. However, the increasing size and complexity of state-of-the-art models present significant challenges for deployment on resource-limited devices. Knowledge distillation (KD) is a promising approach for compressing models, but its effectiveness diminishes when there is a large capacity gap between teacher and student models. To address this issue, we propose Evolving Knowledge Distillation (EKD), a progressive training framework in which the student model learns from a sequence of teachers with gradually increasing capacities. Experiments on IWSLT-14, WMT-17, and WMT-23 benchmarks show that EKD leads to consistent improvements at each stage. On IWSLT-14, the final student achieves a BLEU score of 34.24, narrowing the gap to the strongest teacher (34.32 BLEU) to just 0.08 BLEU. Similar trends are observed on other datasets. These results demonstrate that EKD effectively bridges the capacity gap, enabling compact models to achieve performance close to that of much larger teacher models.Code and models are available at https://github.com/agi-content-generation/EKD.
Training a language model on data scattered across bandwidth-limited nodes that cannot be centralized is a setting that arises in clinical networks, enterprise knowledge bases, and scientific consortia. We study the regime in which data must remain distributed across nodes, and ask what statistical guarantees are in principle achievable under explicit bandwidth budgets; we aim to characterize what is provably possible, not to demonstrate a deployment-ready system. Existing theory treats either training-time consistency or inference-time calibration in isolation, and none makes bandwidth a first-class statistical parameter. We analyze two protocols, Federated Probe-Logit Distillation (FPLD) for training and Federated Conformal RAG (FC-RAG) for inference, as the analytical vehicles for our results. Our first main result is an explicit high-probability KL-consistency rate for FPLD with simultaneous dependence on node count $K$, per-node sample size $n$, quantization budget $B$, probe-set size $m$, and vocabulary size $V$; bandwidth enters only through an exponentially vanishing quantization term. Our second main result is a distribution-free marginal-coverage bound for FC-RAG, whose novel retrieval-bandwidth slack $Δ_{\mathrm{RAG}} = f_{\max}\sqrt{K^{-2}\sum_i v(B_i)}$ makes per-node retrieval bandwidth a first-class statistical parameter, with arithmetic aggregation across $K$ nodes shrinking the slack as $K^{-1/2}$ in the per-node-uniform regime. A Pinsker-type corollary composes the two bounds into an end-to-end coverage guarantee. Synthetic experiments verify the predicted scaling along the bounds' parameters; small-scale experiments on a GPT-2 testbed illustrate that the qualitative bandwidth-accuracy tradeoff survives on a real language model. A deployment-scale empirical evaluation is out of scope.
展开完整摘要收起摘要↓
Training a language model on data scattered across bandwidth-limited nodes that cannot be centralized is a setting that arises in clinical networks, enterprise knowledge bases, and scientific consortia. We study the regime in which data must remain distributed across nodes, and ask what statistical guarantees are in principle achievable under explicit bandwidth budgets; we aim to characterize what is provably possible, not to demonstrate a deployment-ready system. Existing theory treats either training-time consistency or inference-time calibration in isolation, and none makes bandwidth a first-class statistical parameter. We analyze two protocols, Federated Probe-Logit Distillation (FPLD) for training and Federated Conformal RAG (FC-RAG) for inference, as the analytical vehicles for our results. Our first main result is an explicit high-probability KL-consistency rate for FPLD with simultaneous dependence on node count $K$, per-node sample size $n$, quantization budget $B$, probe-set size $m$, and vocabulary size $V$; bandwidth enters only through an exponentially vanishing quantization term. Our second main result is a distribution-free marginal-coverage bound for FC-RAG, whose novel retrieval-bandwidth slack $Δ_{\mathrm{RAG}} = f_{\max}\sqrt{K^{-2}\sum_i v(B_i)}$ makes per-node retrieval bandwidth a first-class statistical parameter, with arithmetic aggregation across $K$ nodes shrinking the slack as $K^{-1/2}$ in the per-node-uniform regime. A Pinsker-type corollary composes the two bounds into an end-to-end coverage guarantee. Synthetic experiments verify the predicted scaling along the bounds' parameters; small-scale experiments on a GPT-2 testbed illustrate that the qualitative bandwidth-accuracy tradeoff survives on a real language model. A deployment-scale empirical evaluation is out of scope.
Erasing specific concepts from text-to-image diffusion models is essential for avoiding the generation of copyrighted and explicit content. Closed-form concept erasure methods offer a fast alternative to backpropagation-based techniques, but they become less effective when scaling from smaller models such as Stable Diffusion 1.5 to larger models like Stable Diffusion XL. To maintain erasure effectiveness in these larger-scale architectures, we propose SParse cross-Attention-based Concept Erasure (SPACE). SPACE iteratively modifies the cross-attention parameters of a model with a closed-form update that jointly induces sparsity and erases target concepts. By concentrating the concept mapping to a lower-dimensional subspace, SPACE achieves superior erasure efficacy compared to dense baselines. Extensive experimental results show improvements in erasure effectiveness and robustness against adversarial prompts. Furthermore, SPACE achieves 80%-90% cross-attention sparsity, reducing the storage requirements for saving the modified parameters by 70%, demonstrating its memory efficiency.
展开完整摘要收起摘要↓
Erasing specific concepts from text-to-image diffusion models is essential for avoiding the generation of copyrighted and explicit content. Closed-form concept erasure methods offer a fast alternative to backpropagation-based techniques, but they become less effective when scaling from smaller models such as Stable Diffusion 1.5 to larger models like Stable Diffusion XL. To maintain erasure effectiveness in these larger-scale architectures, we propose SParse cross-Attention-based Concept Erasure (SPACE). SPACE iteratively modifies the cross-attention parameters of a model with a closed-form update that jointly induces sparsity and erases target concepts. By concentrating the concept mapping to a lower-dimensional subspace, SPACE achieves superior erasure efficacy compared to dense baselines. Extensive experimental results show improvements in erasure effectiveness and robustness against adversarial prompts. Furthermore, SPACE achieves 80%-90% cross-attention sparsity, reducing the storage requirements for saving the modified parameters by 70%, demonstrating its memory efficiency.
作者Wilson Wongso, Lihuan Li, Arian Prabowo, Xiachong Lin, Baiyu Chen, Hao Xue, Flora D. Salim
Generating high-fidelity synthetic GPS trajectories is increasingly important for applications in transportation, urban planning, and what-if scenario simulation, especially as privacy concerns limit access to real-world mobility data. Existing trajectory generation models face a trade-off between efficiency and faithfulness to road network topology: continuous-space methods enable fast generation but ignore the road network, while topology-aware approaches rely on search-based autoregressive decoding that limits generation speed. We propose TrajDLM, a topology-aware trajectory generation framework based on block diffusion language models that bridges this gap. TrajDLM models trajectories as sequences of discrete road segments, combining a block diffusion backbone for efficient denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling to ensure coherent and realistic trajectories. Across three city-scale datasets, TrajDLM achieves strong performance on fine-grained local similarity metrics while being up to $2.8\times$ faster than prior work, and demonstrates strong zero-shot transfer across domains, including unseen transportation modes. These results highlight the effectiveness of block-wise discrete diffusion as a scalable approach to accurate and efficient trajectory generation. Our code is available at https://github.com/cruiseresearchgroup/TrajDLM/
展开完整摘要收起摘要↓
Generating high-fidelity synthetic GPS trajectories is increasingly important for applications in transportation, urban planning, and what-if scenario simulation, especially as privacy concerns limit access to real-world mobility data. Existing trajectory generation models face a trade-off between efficiency and faithfulness to road network topology: continuous-space methods enable fast generation but ignore the road network, while topology-aware approaches rely on search-based autoregressive decoding that limits generation speed. We propose TrajDLM, a topology-aware trajectory generation framework based on block diffusion language models that bridges this gap. TrajDLM models trajectories as sequences of discrete road segments, combining a block diffusion backbone for efficient denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling to ensure coherent and realistic trajectories. Across three city-scale datasets, TrajDLM achieves strong performance on fine-grained local similarity metrics while being up to $2.8\times$ faster than prior work, and demonstrates strong zero-shot transfer across domains, including unseen transportation modes. These results highlight the effectiveness of block-wise discrete diffusion as a scalable approach to accurate and efficient trajectory generation. Our code is available at https://github.com/cruiseresearchgroup/TrajDLM/
作者Byeongchan Kim, Arijit Sehanobish, Avinava Dubey, Min-hwan Oh, Krzysztof Choromanski
We present a new class of efficient attention mechanisms applying universal 3D Relative Positional Encoding (RPE) methods given by arbitrary integrable modulation functions $f$. They lead to the new class of 3D-Transformer models, called RelFlexformers, flexibly integrating those RPEs, and characterized by the $O(L \log L)$ time complexity of the attention computation for the $L$-length input sequences. RelFlexformers builds on the theory of the Non-Uniform Fourier Transform (NU-FFT), naturally generalizing several existing efficient RPE-attention methods from structured settings with tokens homogeneously embedded in unweighted grids into general non-structured heterogeneous scenarios, where tokens' positions are arbitrarily distributed in the corresponding 3D spaces. As such, RelFlexformers can be applied in particular to model point clouds. Our extensive empirical evaluation on a large portfolio of 3D datasets confirms quality improvements provided by the NU-FFT-driven attention modulation techniques in the RelFlexformers.
展开完整摘要收起摘要↓
We present a new class of efficient attention mechanisms applying universal 3D Relative Positional Encoding (RPE) methods given by arbitrary integrable modulation functions $f$. They lead to the new class of 3D-Transformer models, called RelFlexformers, flexibly integrating those RPEs, and characterized by the $O(L \log L)$ time complexity of the attention computation for the $L$-length input sequences. RelFlexformers builds on the theory of the Non-Uniform Fourier Transform (NU-FFT), naturally generalizing several existing efficient RPE-attention methods from structured settings with tokens homogeneously embedded in unweighted grids into general non-structured heterogeneous scenarios, where tokens' positions are arbitrarily distributed in the corresponding 3D spaces. As such, RelFlexformers can be applied in particular to model point clouds. Our extensive empirical evaluation on a large portfolio of 3D datasets confirms quality improvements provided by the NU-FFT-driven attention modulation techniques in the RelFlexformers.
Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call attention drift: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both EAGLE3 drafters and MTP heads, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.
展开完整摘要收起摘要↓
Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call attention drift: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both EAGLE3 drafters and MTP heads, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.