General-purpose agents can plan, use tools, and revise their behavior from feedback, but it remains unclear whether these capabilities transfer from digital environments to embodied manipulation. To investigate this question, we introduce LIBERO-Agent, an agent-native benchmark for evaluating these agents in robot manipulation tasks. Rather than asking agents to submit task-level Python control programs or operate through high-level robot skills, LIBERO-Agent provides an interactive robotic environment where agents can select which observations to inspect, process them with their own tools, and issue native action commands. LIBERO-Agent integrates 200 tasks into a common interaction framework and provides a 30-task primary suite that separates perception, short-horizon execution, and long-horizon composition. Results reveal a pronounced reliability gap: while agents perform well on perception and easy short-horizon tasks, their performance degrades substantially on hard short-horizon and long-horizon tasks. Richer observations improve short-horizon manipulation, while demonstration benefits depend on the agent and format. Among these agents, GPT-6 Astra achieves the strongest overall performance. Further analysis shows its major advantage lies in mechanism interaction, especially when sustained physical contact is needed, while its remaining failures stem from cross-stage interference and geometric errors.
展开完整摘要收起摘要↓
General-purpose agents can plan, use tools, and revise their behavior from feedback, but it remains unclear whether these capabilities transfer from digital environments to embodied manipulation. To investigate this question, we introduce LIBERO-Agent, an agent-native benchmark for evaluating these agents in robot manipulation tasks. Rather than asking agents to submit task-level Python control programs or operate through high-level robot skills, LIBERO-Agent provides an interactive robotic environment where agents can select which observations to inspect, process them with their own tools, and issue native action commands. LIBERO-Agent integrates 200 tasks into a common interaction framework and provides a 30-task primary suite that separates perception, short-horizon execution, and long-horizon composition. Results reveal a pronounced reliability gap: while agents perform well on perception and easy short-horizon tasks, their performance degrades substantially on hard short-horizon and long-horizon tasks. Richer observations improve short-horizon manipulation, while demonstration benefits depend on the agent and format. Among these agents, GPT-6 Astra achieves the strongest overall performance. Further analysis shows its major advantage lies in mechanism interaction, especially when sustained physical contact is needed, while its remaining failures stem from cross-stage interference and geometric errors.
作者Hongbo Zhang, Liuyang Song, Quanquan Li, Daqian Yang, Yan Wen, Zhengtao Yao
An agent that remembers what it did on a web page must decide when two pages count as the same. Memories built on observation similarity merge pages that look alike but behave differently, and GUIs are full of such pages: two tabs of one widget or two rows of one menu answer the same click differently. We define the merge rule as an action-conditioned bisimulation over the empirical predictive state graph a frozen agent fills as it acts. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks under an affordance label. Observation similarity never enters the rule, and nothing is trained. It replaces the merge rule of an existing outcome-value memory, so a closed-loop comparison isolates it. On MiniWoB++ it raises success rate over a memoryless agent, while a control taking identical exploratory detours, the prior successor-representation merge, and the same criterion without action conditioning change nothing.
展开完整摘要收起摘要↓
An agent that remembers what it did on a web page must decide when two pages count as the same. Memories built on observation similarity merge pages that look alike but behave differently, and GUIs are full of such pages: two tabs of one widget or two rows of one menu answer the same click differently. We define the merge rule as an action-conditioned bisimulation over the empirical predictive state graph a frozen agent fills as it acts. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks under an affordance label. Observation similarity never enters the rule, and nothing is trained. It replaces the merge rule of an existing outcome-value memory, so a closed-loop comparison isolates it. On MiniWoB++ it raises success rate over a memoryless agent, while a control taking identical exploratory detours, the prior successor-representation merge, and the same criterion without action conditioning change nothing.
作者Yutong Deng, Qi Song, Xi Guo, Tianming Wang, Lei Bao, Jianping Ge
Camera traps accumulated vast, multidimensional data for wildlife monitoring, yet translating raw media archives into meaningful ecological insights remains highly fragmented. Current research workflows require laboriously stitching together disparate analysis tools and scripts, creating steep programming hurdles and complicating end-to-end spatiotemporal analyses. To overcome this fragmentation, we present CamAgent, an autonomous Large Language Model (LLM) agent framework that integrates camera-trap analytical workflows into a unified intelligent ecosystem. CamAgent interprets natural-language ecological intent, schedules computational routing, and executes specialized tools spanning computer-vision perception (e.g., SpeciesNet), CamtrapDP-compatible data management, detection-corrected occupancy modeling, temporal activity analysis, and species co-occurrence networks. The framework automates multi-stage analytical pipelines while maintaining essential data-quality controls and analytical conventions. Consequently, CamAgent significantly reduces manual programming overhead for conservationists, establishing a transparent, scalable, and fully integrated paradigm for camera-trap ecology. Our project is available at https://anonymous.4open.science/r/artifact72c6f4.
展开完整摘要收起摘要↓
Camera traps accumulated vast, multidimensional data for wildlife monitoring, yet translating raw media archives into meaningful ecological insights remains highly fragmented. Current research workflows require laboriously stitching together disparate analysis tools and scripts, creating steep programming hurdles and complicating end-to-end spatiotemporal analyses. To overcome this fragmentation, we present CamAgent, an autonomous Large Language Model (LLM) agent framework that integrates camera-trap analytical workflows into a unified intelligent ecosystem. CamAgent interprets natural-language ecological intent, schedules computational routing, and executes specialized tools spanning computer-vision perception (e.g., SpeciesNet), CamtrapDP-compatible data management, detection-corrected occupancy modeling, temporal activity analysis, and species co-occurrence networks. The framework automates multi-stage analytical pipelines while maintaining essential data-quality controls and analytical conventions. Consequently, CamAgent significantly reduces manual programming overhead for conservationists, establishing a transparent, scalable, and fully integrated paradigm for camera-trap ecology. Our project is available at https://anonymous.4open.science/r/artifact72c6f4.
作者Shengjie Jin, Hengbo Xu, Zelong Sun, YuJie Guo, Zhiwu Lu
Iterative self-distillation enables LLM agents to learn from successive deployments, offering a path toward recursive self-improvement (RSI). Yet our experiments with existing methods reveal a collapse in deployment performance across cycles, while task performance with privileged information (PI) also declines. We address this collapse by prioritizing informative interaction steps for distillation and preserving PI-conditioned behavior as the student becomes the next teacher. We introduce Retentive and Selective Augmentation for Iterative Self-Distillation (ReSAIL), a plug-in augmentation for iterative PI-based self-distillation. ReSAIL selects interaction steps where PI most strongly changes the teacher's predictions and balances the resulting distillation losses across trajectories. It also regularizes the student's PI-conditioned output distributions toward those of the frozen teacher at selected and unselected steps to preserve PI-conditioned behavior for supervision in the next cycle. On ALFWorld and TextCraft, ReSAIL sustains substantial gains across model scales over three cycles, with an average absolute gain of 22.5% in final-cycle success rates when added to self-distillation baselines. Sensitivity-guided selection of offline data also improves action prediction accuracy for multimodal GUI agents on AITZ. These findings provide the first evidence that a more robust learning mechanism can effectively mitigate performance collapse in iterative agent self-distillation over deployment trajectories.
展开完整摘要收起摘要↓
Iterative self-distillation enables LLM agents to learn from successive deployments, offering a path toward recursive self-improvement (RSI). Yet our experiments with existing methods reveal a collapse in deployment performance across cycles, while task performance with privileged information (PI) also declines. We address this collapse by prioritizing informative interaction steps for distillation and preserving PI-conditioned behavior as the student becomes the next teacher. We introduce Retentive and Selective Augmentation for Iterative Self-Distillation (ReSAIL), a plug-in augmentation for iterative PI-based self-distillation. ReSAIL selects interaction steps where PI most strongly changes the teacher's predictions and balances the resulting distillation losses across trajectories. It also regularizes the student's PI-conditioned output distributions toward those of the frozen teacher at selected and unselected steps to preserve PI-conditioned behavior for supervision in the next cycle. On ALFWorld and TextCraft, ReSAIL sustains substantial gains across model scales over three cycles, with an average absolute gain of 22.5% in final-cycle success rates when added to self-distillation baselines. Sensitivity-guided selection of offline data also improves action prediction accuracy for multimodal GUI agents on AITZ. These findings provide the first evidence that a more robust learning mechanism can effectively mitigate performance collapse in iterative agent self-distillation over deployment trajectories.
Large language models (LLMs) are increasingly combined with knowledge graphs (KGs) to ground reasoning in structured evidence. However, most LLM-based KGQA methods rely on traversing existing graph edges and become unreliable when reasoning paths are broken by missing facts. Alternatives that ask LLMs to generate missing knowledge risk introducing hallucinated evidence. We introduce XoG (eXplore-on-Graph), a framework for multi-hop question answering over incomplete KGs that recovers missing reasoning paths from learned graph structure rather than LLM parametric knowledge. XoG combines type-level entity-relation statistics to identify candidate relations with KG embeddings to retrieve plausible missing entities, using the LLM as a semantic selector and reasoner. These mechanisms are integrated into an iterative planning-exploration-reasoning process. Experiments on WebQSP, CWQ, and the Wikidata-based BRINK benchmark show that XoG remains competitive on complete KGs and consistently outperforms comparable methods without task-specific KGQA training under KG incompleteness. These gains persist across multiple LLM backbones, indicating that stronger LLMs alone do not resolve missing graph evidence. XoG also reduces LLM token consumption by up to 33% compared with a closely related planning-based approach.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly combined with knowledge graphs (KGs) to ground reasoning in structured evidence. However, most LLM-based KGQA methods rely on traversing existing graph edges and become unreliable when reasoning paths are broken by missing facts. Alternatives that ask LLMs to generate missing knowledge risk introducing hallucinated evidence. We introduce XoG (eXplore-on-Graph), a framework for multi-hop question answering over incomplete KGs that recovers missing reasoning paths from learned graph structure rather than LLM parametric knowledge. XoG combines type-level entity-relation statistics to identify candidate relations with KG embeddings to retrieve plausible missing entities, using the LLM as a semantic selector and reasoner. These mechanisms are integrated into an iterative planning-exploration-reasoning process. Experiments on WebQSP, CWQ, and the Wikidata-based BRINK benchmark show that XoG remains competitive on complete KGs and consistently outperforms comparable methods without task-specific KGQA training under KG incompleteness. These gains persist across multiple LLM backbones, indicating that stronger LLMs alone do not resolve missing graph evidence. XoG also reduces LLM token consumption by up to 33% compared with a closely related planning-based approach.
Causal action verifiers gate an agent's state-changing tool calls by checking whether each proposed intervention is identifiable against a committed action-state graph, and they issue a certificate that carries the identification argument and a one-sided lower confidence bound. One such verifier, CIVeX, reports zero false executions on a confounded tool-use benchmark. We red-team it by corrupting only the committed graph. Omitting a single bidirected edge takes it from zero false executions to 15.3% at the benchmark's published confounding strength, with 91% of its executions harmful and utility falling from +2.27 to +0.35. Reversing one arrowhead, so that a mediator is committed as a confounder, gives 48.9% false executions and no correct ones. Every one of these actions carries an internally valid certificate. An attestation step that tests each observationally certified execution against a bounded randomised sample detected both attacks, with 2 false alarms in 555 executions on a truthful graph; refusing what fails the test, or cannot be tested, gave zero false executions in every setting we measured. It does not restore beneficial execution: at the published strength 97.1% of beneficial actions are still never executed, because the same misspecification rejects them before attestation runs. Those rejections carry certificates too, and auditing them works, but its cost scales with the number of rejections rather than the number of executions. Recovering safety costs 127 experiments per 1,050 actions; recovering the lost value costs 614 more, at which point the audited verifier makes the honest graph's decisions on every instance and spends exactly its experiment budget. An audit that inspects only executions protects against wrongful action. Wrongful inaction has to be paid for separately.
展开完整摘要收起摘要↓
Causal action verifiers gate an agent's state-changing tool calls by checking whether each proposed intervention is identifiable against a committed action-state graph, and they issue a certificate that carries the identification argument and a one-sided lower confidence bound. One such verifier, CIVeX, reports zero false executions on a confounded tool-use benchmark. We red-team it by corrupting only the committed graph. Omitting a single bidirected edge takes it from zero false executions to 15.3% at the benchmark's published confounding strength, with 91% of its executions harmful and utility falling from +2.27 to +0.35. Reversing one arrowhead, so that a mediator is committed as a confounder, gives 48.9% false executions and no correct ones. Every one of these actions carries an internally valid certificate. An attestation step that tests each observationally certified execution against a bounded randomised sample detected both attacks, with 2 false alarms in 555 executions on a truthful graph; refusing what fails the test, or cannot be tested, gave zero false executions in every setting we measured. It does not restore beneficial execution: at the published strength 97.1% of beneficial actions are still never executed, because the same misspecification rejects them before attestation runs. Those rejections carry certificates too, and auditing them works, but its cost scales with the number of rejections rather than the number of executions. Recovering safety costs 127 experiments per 1,050 actions; recovering the lost value costs 614 more, at which point the audited verifier makes the honest graph's decisions on every instance and spends exactly its experiment budget. An audit that inspects only executions protects against wrongful action. Wrongful inaction has to be paid for separately.
This paper presents a scalable optimal-transport-based framework for terminal distribution matching in multi-agent systems. While optimal transport provides a natural way to measure distributional mismatch and assign agents to a desired spatial distribution, global discrete transport can become computationally expensive for large-scale systems. We address this bottleneck by partitioning agents and target samples into spatially corresponding blocks and solving smaller local transport problems. Under a mass-balance condition, the resulting restricted coupling remains feasible for the global problem and provides an upper bound on the Wasserstein cost. The local assignments generate target locations for finite-horizon agent control, applicable to both linear and nonlinear dynamics. By alternating local assignment and control, we establish a cycle-to-cycle descent guarantee for the resulting transport surrogate. The proposed framework therefore enables scalable terminal distribution matching while retaining a rigorous connection to the Wasserstein objective. The technical soundness of the proposed results is validated through simulations.
展开完整摘要收起摘要↓
This paper presents a scalable optimal-transport-based framework for terminal distribution matching in multi-agent systems. While optimal transport provides a natural way to measure distributional mismatch and assign agents to a desired spatial distribution, global discrete transport can become computationally expensive for large-scale systems. We address this bottleneck by partitioning agents and target samples into spatially corresponding blocks and solving smaller local transport problems. Under a mass-balance condition, the resulting restricted coupling remains feasible for the global problem and provides an upper bound on the Wasserstein cost. The local assignments generate target locations for finite-horizon agent control, applicable to both linear and nonlinear dynamics. By alternating local assignment and control, we establish a cycle-to-cycle descent guarantee for the resulting transport surrogate. The proposed framework therefore enables scalable terminal distribution matching while retaining a rigorous connection to the Wasserstein objective. The technical soundness of the proposed results is validated through simulations.
Conventional image forgery detection methods produce binary scores or pixel-level masks without interpretable evidence, while recent multimodal large language model (MLLM)-based approaches generate post-hoc explanations of predetermined classification results rather than reasoning from evidence. Inspired by the forensic workflow of human judicial experts, we propose Agentic Tool-Augmented Reasoning (ATAR), a framework integrating 22 specialized forensic tools across seven complementary domains to autonomously detect, localize, and explain image forgeries through multi-turn reasoning. A Dual-Stream Forensic Reasoning paradigm combines a high-level semantic anomaly path, which magnifies suspicious regions for fine-grained inspection, with a low-level forgery artifact path, which invokes forensic tools to extract objective evidence. We further introduce Forensics Curriculum Learning: during General Experience SFT, an automated teacher-student mentoring pipeline synthesizes multi-turn tool-usage reasoning trajectories; during Forensic Scene RL, a Tool Prior Curriculum guides early tool exploration and progressively transfers control to the agent, while a Structured Evidence Reward provides fine-grained process-level supervision. Experiments across IMDL, Deepfake detection, DMDL, and AIGC detection show that ATAR achieves 78.5% average image-level F1 on six zero-shot IMDL benchmarks, surpassing the strongest MLLM baseline by 11.8 percentage points, and remains competitive with specialized detectors on other tasks while producing substantially more faithful and grounded explanations.
展开完整摘要收起摘要↓
Conventional image forgery detection methods produce binary scores or pixel-level masks without interpretable evidence, while recent multimodal large language model (MLLM)-based approaches generate post-hoc explanations of predetermined classification results rather than reasoning from evidence. Inspired by the forensic workflow of human judicial experts, we propose Agentic Tool-Augmented Reasoning (ATAR), a framework integrating 22 specialized forensic tools across seven complementary domains to autonomously detect, localize, and explain image forgeries through multi-turn reasoning. A Dual-Stream Forensic Reasoning paradigm combines a high-level semantic anomaly path, which magnifies suspicious regions for fine-grained inspection, with a low-level forgery artifact path, which invokes forensic tools to extract objective evidence. We further introduce Forensics Curriculum Learning: during General Experience SFT, an automated teacher-student mentoring pipeline synthesizes multi-turn tool-usage reasoning trajectories; during Forensic Scene RL, a Tool Prior Curriculum guides early tool exploration and progressively transfers control to the agent, while a Structured Evidence Reward provides fine-grained process-level supervision. Experiments across IMDL, Deepfake detection, DMDL, and AIGC detection show that ATAR achieves 78.5% average image-level F1 on six zero-shot IMDL benchmarks, surpassing the strongest MLLM baseline by 11.8 percentage points, and remains competitive with specialized detectors on other tasks while producing substantially more faithful and grounded explanations.
作者Jihun Han, Yejin Jang, Byung Il Kwak, Mee Lan Han
LLM-based agents extend their capabilities through third-party skills that provide task-specific instructions, scripts, and tool-use procedures. However, malicious instructions inserted into an otherwise benign skill can cause a benign user request to trigger dangerous Tool Calls, including data exfiltration, file deletion, or unauthorized code execution. This paper presents ActionGuard, which inspects skill-influenced Tool Calls immediately before execution. ActionGuard separates the target agent's action-generation context from the safeguard's authorization context. The target agent may use the original skill for planning, but the Reviewer does not receive the potentially poisoned raw skill text. Instead, it determines whether each action is justified by the trusted user request using a balanced skill profile, current and recent Tool Calls, and local script contents. ActionGuard intercepts each Tool Call at OpenClaw's before-tool-call stage and enforces the Reviewer's ALLOW or DENY decision under a fail-closed policy. We evaluate ActionGuard on 139 contextual and 180 obvious injections in a SKILL-INJECT-based setting against Dynamic Guardian and SkillGuard, using three open-source and two commercial Reviewer models. Each condition is repeated three times and evaluated using Attack Success Rate (ASR) and Task Success Rate (TSR). Overall, ActionGuard reduced ASR by 35.54 to 46.11 percent relative to existing safeguards and by 70.44 percent relative to No Safeguard, while maintaining high benign-task completion. These results show that execution-boundary authorization grounded in trusted user intent and runtime evidence can restrict unauthorized Tool Calls induced by skill injection.
展开完整摘要收起摘要↓
LLM-based agents extend their capabilities through third-party skills that provide task-specific instructions, scripts, and tool-use procedures. However, malicious instructions inserted into an otherwise benign skill can cause a benign user request to trigger dangerous Tool Calls, including data exfiltration, file deletion, or unauthorized code execution. This paper presents ActionGuard, which inspects skill-influenced Tool Calls immediately before execution. ActionGuard separates the target agent's action-generation context from the safeguard's authorization context. The target agent may use the original skill for planning, but the Reviewer does not receive the potentially poisoned raw skill text. Instead, it determines whether each action is justified by the trusted user request using a balanced skill profile, current and recent Tool Calls, and local script contents. ActionGuard intercepts each Tool Call at OpenClaw's before-tool-call stage and enforces the Reviewer's ALLOW or DENY decision under a fail-closed policy. We evaluate ActionGuard on 139 contextual and 180 obvious injections in a SKILL-INJECT-based setting against Dynamic Guardian and SkillGuard, using three open-source and two commercial Reviewer models. Each condition is repeated three times and evaluated using Attack Success Rate (ASR) and Task Success Rate (TSR). Overall, ActionGuard reduced ASR by 35.54 to 46.11 percent relative to existing safeguards and by 70.44 percent relative to No Safeguard, while maintaining high benign-task completion. These results show that execution-boundary authorization grounded in trusted user intent and runtime evidence can restrict unauthorized Tool Calls induced by skill injection.
LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline---combining supervised fine-tuning, online imitation learning, and reinforcement learning---that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5% to 76.3%. In agentic search, it improves task success rates by 52.7% and reduces the average token length by 61% from 1886 to 728, whereas upgrading the agent yields only a 3.7% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.
展开完整摘要收起摘要↓
LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline---combining supervised fine-tuning, online imitation learning, and reinforcement learning---that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5% to 76.3%. In agentic search, it improves task success rates by 52.7% and reduces the average token length by 61% from 1886 to 728, whereas upgrading the agent yields only a 3.7% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.
As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.
展开完整摘要收起摘要↓
As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.
作者Bo Yin, Dongbo Li, Hongkai Chen, Jie Liu, Guoliang Xing
Constrained multi-agent control requires more than predicting rewarding actions: an action can cease to be executable as contact windows, shared capacity, and deadlines change. We introduce VERA, a centralized-training, decentralized-execution framework that separates feasibility estimation from credit assignment. Each actor predicts a five-dimensional verifiable feasibility representation (VFR). After an action is proposed, exact action-conditioned margins available only during training supervise that representation, while a counterfactual group-relative advantage (CGRA) ranks candidate representation-action pairs. Execution uses one actor pass and no privileged state. In a dynamic space-air-ground integrated network (SAGIN), VERA obtains 55.33% +/- 3.60% success with 0.45% +/- 0.81% coverage violation, within 1.33 percentage points of a privileged-mask reference. With rewards matched over ten paired seeds, VERA improves success over the strongest baseline by 8.74 percentage points (p=0.023) and reduces violation by 52.19 percentage points (p=5.7e-8). A ten-seed 4-by-2 factorial attributes a 14.16-16.48 percentage-point gain to CGRA across handcrafted, learned, random, and latent representations; evaluation on seven unseen topologies preserves a 24.33-30.02 percentage-point advantage over multi-agent proximal policy optimization. From 10 to 40 users, success remains 50.1-53.8%, and VFR adds only 0.026 ms to a central processing unit (CPU) actor step. Cross-domain tests further identify the governing condition: counterfactual credit succeeds when candidate scores respect shared constraints and fails under incompatible reward geometries. These results establish action-conditioned feasibility as an auditable training interface and counterfactual credit as a geometry-dependent optimization mechanism.
展开完整摘要收起摘要↓
Constrained multi-agent control requires more than predicting rewarding actions: an action can cease to be executable as contact windows, shared capacity, and deadlines change. We introduce VERA, a centralized-training, decentralized-execution framework that separates feasibility estimation from credit assignment. Each actor predicts a five-dimensional verifiable feasibility representation (VFR). After an action is proposed, exact action-conditioned margins available only during training supervise that representation, while a counterfactual group-relative advantage (CGRA) ranks candidate representation-action pairs. Execution uses one actor pass and no privileged state. In a dynamic space-air-ground integrated network (SAGIN), VERA obtains 55.33% +/- 3.60% success with 0.45% +/- 0.81% coverage violation, within 1.33 percentage points of a privileged-mask reference. With rewards matched over ten paired seeds, VERA improves success over the strongest baseline by 8.74 percentage points (p=0.023) and reduces violation by 52.19 percentage points (p=5.7e-8). A ten-seed 4-by-2 factorial attributes a 14.16-16.48 percentage-point gain to CGRA across handcrafted, learned, random, and latent representations; evaluation on seven unseen topologies preserves a 24.33-30.02 percentage-point advantage over multi-agent proximal policy optimization. From 10 to 40 users, success remains 50.1-53.8%, and VFR adds only 0.026 ms to a central processing unit (CPU) actor step. Cross-domain tests further identify the governing condition: counterfactual credit succeeds when candidate scores respect shared constraints and fails under incompatible reward geometries. These results establish action-conditioned feasibility as an auditable training interface and counterfactual credit as a geometry-dependent optimization mechanism.
Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent
展开完整摘要收起摘要↓
Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent
作者Shuo Zhang, Yifan Zhou, Han Wang, Jinsong Zhang, Jingyu Li, Hongbing Li, Zhejun Zhang, Chengyi Zhao, Yuquan Hao, Yitong Liu, Jiyin Li, Ruiqi Tang, Zixuan Lin, Yi Luo, Xurui Zhang, Ronghao Chen, Huacan Wang, Lei Li
While recent Multimodal Large Language Models (MLLMs) have shown promise in affective computing, their reasoning capabilities are largely confined to short video clips with limited interactions. However, real-world emotions are not merely isolated instantaneous reactions but dynamic and cumulative processes deeply shaped by past experiences and ongoing events. To bridge this gap, we introduce LongEmoBench, a benchmark dedicated to emotion understanding and reasoning in long videos. It assesses progressive capabilities scaling from continuous scene interactions to complex episodic developments. Furthermore, we propose LongEmo, a novel memory-augmented agentic framework designed to tackle the immense challenges of long-range affective reasoning. LongEmo processes continuous video streams to construct an Event Memory Graph, explicitly modeling long-range dependencies and capturing emotional dynamics across discrete events. Given a question, the agent retrieves a query-relevant event stream from the graph, iteratively integrating multimodal memories and relational dependencies to deduce the final answer. Extensive evaluations of 17 representative methods reveal that they struggle significantly with emotion understanding and reasoning in long videos. In contrast, LongEmo achieves state-of-the-art performance, demonstrating the efficacy of its event-centric memory architecture.
展开完整摘要收起摘要↓
While recent Multimodal Large Language Models (MLLMs) have shown promise in affective computing, their reasoning capabilities are largely confined to short video clips with limited interactions. However, real-world emotions are not merely isolated instantaneous reactions but dynamic and cumulative processes deeply shaped by past experiences and ongoing events. To bridge this gap, we introduce LongEmoBench, a benchmark dedicated to emotion understanding and reasoning in long videos. It assesses progressive capabilities scaling from continuous scene interactions to complex episodic developments. Furthermore, we propose LongEmo, a novel memory-augmented agentic framework designed to tackle the immense challenges of long-range affective reasoning. LongEmo processes continuous video streams to construct an Event Memory Graph, explicitly modeling long-range dependencies and capturing emotional dynamics across discrete events. Given a question, the agent retrieves a query-relevant event stream from the graph, iteratively integrating multimodal memories and relational dependencies to deduce the final answer. Extensive evaluations of 17 representative methods reveal that they struggle significantly with emotion understanding and reasoning in long videos. In contrast, LongEmo achieves state-of-the-art performance, demonstrating the efficacy of its event-centric memory architecture.
Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. It implements these strategies and evaluates their combinations to improve task performance. Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. Experiments demonstrate improvements on AppWorld and Tau2-Bench. ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom.
展开完整摘要收起摘要↓
Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. It implements these strategies and evaluates their combinations to improve task performance. Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. Experiments demonstrate improvements on AppWorld and Tau2-Bench. ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom.
作者Yang Cai, Vineet Gupta, Yanchen Jiang, Christopher Liaw, Aranyak Mehta, Grigoris Velegkas, Di Wang
We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Cogentic addresses these challenges through an iterative prove-verify loop in which an orchestrator allocates a population of independent provers across distinct proof directions, subjects their output to adversarial verification by several specialized components, and promotes confirmed intermediate results into a persistent verified ledger that later rounds build on. The harness is designed to be able to solve research-level math and theoretical computer science problems. Using either Gemini 3.1 Pro or an early version of Gemini 4 Argon as the base model, Cogentic produced novel results on open problems across online learning, auction theory, and mechanism design. Each result was independently verified by domain experts and is developed in full in companion papers. We list these results, and new ones as they are verified, at https://sites.google.com/view/cogentic .
展开完整摘要收起摘要↓
We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Cogentic addresses these challenges through an iterative prove-verify loop in which an orchestrator allocates a population of independent provers across distinct proof directions, subjects their output to adversarial verification by several specialized components, and promotes confirmed intermediate results into a persistent verified ledger that later rounds build on. The harness is designed to be able to solve research-level math and theoretical computer science problems. Using either Gemini 3.1 Pro or an early version of Gemini 4 Argon as the base model, Cogentic produced novel results on open problems across online learning, auction theory, and mechanism design. Each result was independently verified by domain experts and is developed in full in companion papers. We list these results, and new ones as they are verified, at https://sites.google.com/view/cogentic .
While large language models have achieved remarkable success in isolated code generation, authentic software engineering requires sustained reasoning, complex state management, and continuous cross-domain abstraction. However, current evaluations of skill evolution in autonomous agents suffer from a critical identifiability problem: they structurally confound genuine capability abstraction with rote solution leakage (i.e., copying highly similar code from historical training data). To resolve this, we introduce EngramBench, a rigorous, capability-grounded benchmark governed by the strict axiom of capability overlap without solution overlap. Comprising 30 diverse learning tasks and 13 unseen transfer tasks, EngramBench challenges agents to navigate interactive, multi-hour development cycles driven by LLM-simulated users. Our extensive evaluation across 48 multi-hour execution trajectories — corroborated by human-expert validation — reveals a profound insight into procedural memory. We demonstrate that static skill banks do not magically bypass the "last mile" of exact code implementation, which remains bottlenecked by the base model's inherent reasoning limits. However, they serve as an indispensable execution compass. By navigating agents away from catastrophic, token-heavy trial-and-error, genuine capability abstraction slashes redundant context bloat and reduces overall coding time by over 55%. Ultimately, EngramBench shifts the evaluation paradigm from trivial pattern matching to the verifiable measurement of deep, cross-domain capability transfer.
展开完整摘要收起摘要↓
While large language models have achieved remarkable success in isolated code generation, authentic software engineering requires sustained reasoning, complex state management, and continuous cross-domain abstraction. However, current evaluations of skill evolution in autonomous agents suffer from a critical identifiability problem: they structurally confound genuine capability abstraction with rote solution leakage (i.e., copying highly similar code from historical training data). To resolve this, we introduce EngramBench, a rigorous, capability-grounded benchmark governed by the strict axiom of capability overlap without solution overlap. Comprising 30 diverse learning tasks and 13 unseen transfer tasks, EngramBench challenges agents to navigate interactive, multi-hour development cycles driven by LLM-simulated users. Our extensive evaluation across 48 multi-hour execution trajectories — corroborated by human-expert validation — reveals a profound insight into procedural memory. We demonstrate that static skill banks do not magically bypass the "last mile" of exact code implementation, which remains bottlenecked by the base model's inherent reasoning limits. However, they serve as an indispensable execution compass. By navigating agents away from catastrophic, token-heavy trial-and-error, genuine capability abstraction slashes redundant context bloat and reduces overall coding time by over 55%. Ultimately, EngramBench shifts the evaluation paradigm from trivial pattern matching to the verifiable measurement of deep, cross-domain capability transfer.
作者Jingwei Jia, Keyu Zhou, Jiewei Wang, Peisen Xu, Xingyuan Zhou, Liang Wang, Jiming Chen, Gaofeng Li, Jin Wang, Shunlei Li
Long-horizon surgical assistance requires humanoid robots to coordinate with evolving human activities while maintaining safety across planning and execution. We present RoboAssist, an agent-based framework for interactive human-humanoid planning that integrates workflow reasoning, task coordination, and cross-layer safety. At its core is an asymmetric dual-track representation that separates partially observed human process states from executable robot task sequences. By updating human-process estimates, scene context, and task dependencies online, RoboAssist revalidates the remaining task sequence and replans only the affected suffix when workflow requests change. A cross-layer safety architecture combines preventive navigation regulation, reactive regulation during close-range handover, and independent whole-body runtime supervision. This design couples online task coordination with safety constraints throughout execution. We demonstrate the framework on a Unitree G1 humanoid robot in long-horizon, multi-stage simulated surgical assistance scenarios encompassing multimodal interaction, instrument handling, medical material transport, navigation, and safe human-robot handover. Experiments show multi-stage task completion and adaptation to workflow-request changes. A targeted full-replanning ablation shows that residual replanning reduces plan-update latency and post-update token usage. Separate safety experiments demonstrate complementary protection across navigation, handover, and runtime supervision. Additional results and demonstrations are available online at https://roboassist.github.io.
展开完整摘要收起摘要↓
Long-horizon surgical assistance requires humanoid robots to coordinate with evolving human activities while maintaining safety across planning and execution. We present RoboAssist, an agent-based framework for interactive human-humanoid planning that integrates workflow reasoning, task coordination, and cross-layer safety. At its core is an asymmetric dual-track representation that separates partially observed human process states from executable robot task sequences. By updating human-process estimates, scene context, and task dependencies online, RoboAssist revalidates the remaining task sequence and replans only the affected suffix when workflow requests change. A cross-layer safety architecture combines preventive navigation regulation, reactive regulation during close-range handover, and independent whole-body runtime supervision. This design couples online task coordination with safety constraints throughout execution. We demonstrate the framework on a Unitree G1 humanoid robot in long-horizon, multi-stage simulated surgical assistance scenarios encompassing multimodal interaction, instrument handling, medical material transport, navigation, and safe human-robot handover. Experiments show multi-stage task completion and adaptation to workflow-request changes. A targeted full-replanning ablation shows that residual replanning reduces plan-update latency and post-update token usage. Separate safety experiments demonstrate complementary protection across navigation, handover, and runtime supervision. Additional results and demonstrations are available online at https://roboassist.github.io.
作者Quan M. Tran, Zhuo Huang, Zhen Fang, Jing Zhang, Mingming Gong, Tongliang Liu
Autonomous agents increasingly rely on memory to generalize beyond their training environments. However, agents are bounded by what they have seen and believed, and leveraging such memories in unseen environments can introduce biases into their internal beliefs. We formalize this phenomenon as false memory, which can arise from spurious correlations, environment shifts, and knowledge conflicts. Despite its importance, false memory is difficult to evaluate because it stems from agent internal beliefs and is easily confounded with ordinary generalization failures. Therefore, we propose FAME, a training-free framework that evaluates false memory through the evolution of agent beliefs under counterfactual reasoning. Specifically, counterfactual scenarios reveal how beliefs change as the latent concept of memory shifts under hypothetical interventions; thus, measuring the resulting concept drift provides a signal for distinguishing faithful versus false memory. Such concepts can be estimated from agent hidden states before answer generation, avoiding the need for reward design or answer sampling. Empirical experiments reveal that simply monitoring answers often fails to detect false memory, while FAME achieves AUROCs of 76.2% - 96.7% across false-memory settings, and outperforms the best baseline by 3.4% - 23.3% across realistic benchmarks, spanning math reasoning (GSM-Symbolic), code generation (GitChameleon), and complex reasoning (BigBench-Hard). We further release corresponding counterfactual templates and facilitate future research on false memory.
展开完整摘要收起摘要↓
Autonomous agents increasingly rely on memory to generalize beyond their training environments. However, agents are bounded by what they have seen and believed, and leveraging such memories in unseen environments can introduce biases into their internal beliefs. We formalize this phenomenon as false memory, which can arise from spurious correlations, environment shifts, and knowledge conflicts. Despite its importance, false memory is difficult to evaluate because it stems from agent internal beliefs and is easily confounded with ordinary generalization failures. Therefore, we propose FAME, a training-free framework that evaluates false memory through the evolution of agent beliefs under counterfactual reasoning. Specifically, counterfactual scenarios reveal how beliefs change as the latent concept of memory shifts under hypothetical interventions; thus, measuring the resulting concept drift provides a signal for distinguishing faithful versus false memory. Such concepts can be estimated from agent hidden states before answer generation, avoiding the need for reward design or answer sampling. Empirical experiments reveal that simply monitoring answers often fails to detect false memory, while FAME achieves AUROCs of 76.2% - 96.7% across false-memory settings, and outperforms the best baseline by 3.4% - 23.3% across realistic benchmarks, spanning math reasoning (GSM-Symbolic), code generation (GitChameleon), and complex reasoning (BigBench-Hard). We further release corresponding counterfactual templates and facilitate future research on false memory.
作者Huaiyu Fu, Heng Cao, Hao Wang, Jian Ya, Tao Chen
LLM agents often admit multiple high-quality solutions to the same task, differing in reasoning structure, tool-use pattern, or interaction trajectory. Yet existing notions of diversity in LLM post-training are mostly implicit, arising from general stochasticity and regularization mechanisms rather than explicitly targeting task-relevant behavioral variation. While such implicit diversity can be useful, it does not directly specify which forms of behavioral variation should be encouraged for a given task. In this work, we study explicit trajectory diversity in RL-based post-training for LLMs. Our key idea is to define diversity through user-specified, task-specific trajectory descriptors, which map each sampled trajectory to an interpretable behavioral representation, and then measure diversity as a set-level functional over the resulting descriptor matrix. Building on this formulation, we introduce Trajectory-guided Joint Policy Optimization(TJPO), a single-policy framework that optimizes explicit diversity over sampled trajectory groups, avoiding the need for population-based policy training, and instantiate it within group-based policy optimization through trajectory-level learning signals. This design makes the diversity objective both interpretable and controllable. Experiments on Sokoban and ALFWorld show that TJPO improves task-specific trajectory diversity while maintaining competitive task performance. Descriptor and trajectory analyses show that the learned variation follows the specified behavioral dimensions and includes distinct successful strategies. Extra experiment results suggest that explicitly shaping trajectory diversity can help LLM agents satisfy user requirements and remain effective when task conditions change.
展开完整摘要收起摘要↓
LLM agents often admit multiple high-quality solutions to the same task, differing in reasoning structure, tool-use pattern, or interaction trajectory. Yet existing notions of diversity in LLM post-training are mostly implicit, arising from general stochasticity and regularization mechanisms rather than explicitly targeting task-relevant behavioral variation. While such implicit diversity can be useful, it does not directly specify which forms of behavioral variation should be encouraged for a given task. In this work, we study explicit trajectory diversity in RL-based post-training for LLMs. Our key idea is to define diversity through user-specified, task-specific trajectory descriptors, which map each sampled trajectory to an interpretable behavioral representation, and then measure diversity as a set-level functional over the resulting descriptor matrix. Building on this formulation, we introduce Trajectory-guided Joint Policy Optimization(TJPO), a single-policy framework that optimizes explicit diversity over sampled trajectory groups, avoiding the need for population-based policy training, and instantiate it within group-based policy optimization through trajectory-level learning signals. This design makes the diversity objective both interpretable and controllable. Experiments on Sokoban and ALFWorld show that TJPO improves task-specific trajectory diversity while maintaining competitive task performance. Descriptor and trajectory analyses show that the learned variation follows the specified behavioral dimensions and includes distinct successful strategies. Extra experiment results suggest that explicitly shaping trajectory diversity can help LLM agents satisfy user requirements and remain effective when task conditions change.
Optimization modeling formulates real-world decision problems as mathematical programs that solvers can use to find optimal decisions. Large language models (LLMs) can automate this process, but the resulting correct formulations can require substantial time and memory to construct and solve, limiting practical scalability. Therefore, we systematically investigate whether LLMs can identify problem structure from natural-language descriptions and apply suitable optimization modeling techniques to generate mathematical models and solver code that solve the problems correctly and efficiently. To this end, we first curate OptTips, a knowledge base of 50 expert modeling techniques in eight families. Using this knowledge, we develop OptDachshund, a multi-agent framework that transforms problems from existing optimization benchmarks into new tasks for evaluating LLMs' use of modeling techniques. It constructs conventional and expert mathematical models with solver code for the same task and data, providing baselines for correctness and computational cost. The resulting EfficientOpt benchmark contains 561 expert-reviewed tasks with paired reference implementations. Evaluation of 11 representative LLMs reveals an efficiency gap on correctly solved tasks with comparable measurements: for every LLM, most generated programs take longer to solve than their expert counterparts. Within the comparable reference-size subset, 57% of programs with correct objective values and fewer variables and linear constraints have longer recorded solver times. Case studies show that different modeling techniques can achieve the same optimal value at similar recorded cost. Faster solving may not reduce execution time if the code takes longer to prepare data and build the model. LLM optimization modeling should therefore be evaluated for both correctness and computational efficiency.
展开完整摘要收起摘要↓
Optimization modeling formulates real-world decision problems as mathematical programs that solvers can use to find optimal decisions. Large language models (LLMs) can automate this process, but the resulting correct formulations can require substantial time and memory to construct and solve, limiting practical scalability. Therefore, we systematically investigate whether LLMs can identify problem structure from natural-language descriptions and apply suitable optimization modeling techniques to generate mathematical models and solver code that solve the problems correctly and efficiently. To this end, we first curate OptTips, a knowledge base of 50 expert modeling techniques in eight families. Using this knowledge, we develop OptDachshund, a multi-agent framework that transforms problems from existing optimization benchmarks into new tasks for evaluating LLMs' use of modeling techniques. It constructs conventional and expert mathematical models with solver code for the same task and data, providing baselines for correctness and computational cost. The resulting EfficientOpt benchmark contains 561 expert-reviewed tasks with paired reference implementations. Evaluation of 11 representative LLMs reveals an efficiency gap on correctly solved tasks with comparable measurements: for every LLM, most generated programs take longer to solve than their expert counterparts. Within the comparable reference-size subset, 57% of programs with correct objective values and fewer variables and linear constraints have longer recorded solver times. Case studies show that different modeling techniques can achieve the same optimal value at similar recorded cost. Faster solving may not reduce execution time if the code takes longer to prepare data and build the model. LLM optimization modeling should therefore be evaluated for both correctness and computational efficiency.
作者Ziyan Jiang, Jingbo Yang, Jiabao Ji, Yujian Liu, Qiucheng Wu, Tommi Jaakkola, Yang Zhang, Shiyu Chang
As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or objects inconsistent with the surrounding scene. Multimodal AI systems, including vision-language models (VLMs) and vision-language-action models (VLAs), have shown potential for automating this task. However, 3D world auditing is complex, requiring the close coupling of two distinct capabilities: action, to navigate the 3D world and search for anomalies systematically and efficiently; and visual reasoning, to understand the environment and identify anomalies from multimodal observations. It remains largely unexplored whether multimodal agents can effectively couple these two capabilities, using visual reasoning to identify potential anomalies while taking actions to validate them. In this paper, we introduce WorldAuditBench, a benchmark for 3D world auditing comprising 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, spanning five anomaly families. We evaluate five frontier models under a fixed exploration budget using two auditing paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection. Across the evaluated models and two paradigms, success rates range from 6.6% to 42.3%, substantially below human performance (83.4%). Through the task of world auditing, WorldAuditBench provides a testbed for studying how multimodal agents couple action and visual reasoning in interactive 3D environments, while highlighting current limitations in their ability to gather and interpret evidence during exploration.
展开完整摘要收起摘要↓
As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or objects inconsistent with the surrounding scene. Multimodal AI systems, including vision-language models (VLMs) and vision-language-action models (VLAs), have shown potential for automating this task. However, 3D world auditing is complex, requiring the close coupling of two distinct capabilities: action, to navigate the 3D world and search for anomalies systematically and efficiently; and visual reasoning, to understand the environment and identify anomalies from multimodal observations. It remains largely unexplored whether multimodal agents can effectively couple these two capabilities, using visual reasoning to identify potential anomalies while taking actions to validate them. In this paper, we introduce WorldAuditBench, a benchmark for 3D world auditing comprising 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, spanning five anomaly families. We evaluate five frontier models under a fixed exploration budget using two auditing paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection. Across the evaluated models and two paradigms, success rates range from 6.6% to 42.3%, substantially below human performance (83.4%). Through the task of world auditing, WorldAuditBench provides a testbed for studying how multimodal agents couple action and visual reasoning in interactive 3D environments, while highlighting current limitations in their ability to gather and interpret evidence during exploration.
GUI agents built on large language and vision-language models still struggle on unseen applications and complex multi-step tasks, as completing real GUI tasks depends on app-specific, temporally volatile operational knowledge that is scarce in pretraining corpora. Retrieval-augmented execution offers a natural remedy but faces two coupled bottlenecks: knowledge at scale is hard to acquire, and self-collected priors inevitably drift from the live environment due to version updates, promotions, ads, A/B tests, and personalization. We therefore argue that GUI agents should not pursue perfect knowledge but learn to act correctly under imperfect priors, and propose our framework that couples knowledge acquisition with noise-robust utilization: a structured exploration strategy traverses interactive elements, builds a UI state-transition graph, and synthesizes (task, trajectory) pairs via a VLM without human annotation; a noise-aware training strategy, grounded in a taxonomy of real GUI drift patterns, injects five types of realistic errors into self-explored trajectories to teach the agent to assess prior reliability before acting. Experiments on physical devices and online emulator benchmarks show that our method discovers more unique screens, covers more benchmark tasks, and more effectively rejects erroneous priors while leveraging correct ones, with accuracy gains that transfer across datasets.
展开完整摘要收起摘要↓
GUI agents built on large language and vision-language models still struggle on unseen applications and complex multi-step tasks, as completing real GUI tasks depends on app-specific, temporally volatile operational knowledge that is scarce in pretraining corpora. Retrieval-augmented execution offers a natural remedy but faces two coupled bottlenecks: knowledge at scale is hard to acquire, and self-collected priors inevitably drift from the live environment due to version updates, promotions, ads, A/B tests, and personalization. We therefore argue that GUI agents should not pursue perfect knowledge but learn to act correctly under imperfect priors, and propose our framework that couples knowledge acquisition with noise-robust utilization: a structured exploration strategy traverses interactive elements, builds a UI state-transition graph, and synthesizes (task, trajectory) pairs via a VLM without human annotation; a noise-aware training strategy, grounded in a taxonomy of real GUI drift patterns, injects five types of realistic errors into self-explored trajectories to teach the agent to assess prior reliability before acting. Experiments on physical devices and online emulator benchmarks show that our method discovers more unique screens, covers more benchmark tasks, and more effectively rejects erroneous priors while leveraging correct ones, with accuracy gains that transfer across datasets.
Large language models (LLMs) are increasingly deployed as automated judges for AI-generated content, yet a single judge is unreliable and even a panel of judges leaves a hard residue: when judges disagree, majority voting discards the conflict instead of resolving it. We present JuryFlow, a disagreement-guided, human-in-the-loop multi-agent evaluation framework that treats inter-judge disagreement not as noise to be averaged away, but as a precise, claim-level signal indicating where an evaluation is uncertain. JuryFlow decomposes each candidate response into atomic claims, has a panel of heterogeneous judges assign per-claim verdicts, and builds a disagreement graph whose nodes are scored by verdict entropy and whose edges encode structural similarity between claims. A human acts as a structural guide, selecting which disagreement to resolve through a single, minimal intervention rather than re-labeling the response, after which the focal claim is re-evaluated, the correction propagates along graph edges and to historically similar cases, and is crystallized into reusable rubric entries that all judges inherit, making the evaluator progressively self-refining. To enable large-scale, reproducible benchmarking without human studies, we evaluate JuryFlow in an automatic configuration in which focal selection is made by entropy ranking. On MT-Bench and LLMBar, JuryFlow improves agreement with gold labels over single-judge and majority-vote panel baselines, and ablations isolate the contributions of disagreement-targeted re-evaluation, propagation, and rubric induction. We contribute (1) a human-in-the-loop paradigm that recasts the human from labeler to structural guide, (2) the JuryFlow framework operationalizing it through a disagreement graph, focal re-evaluation, and closed-loop rubric induction, and (3) an evaluation protocol with ablations that isolate where the gains originate.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly deployed as automated judges for AI-generated content, yet a single judge is unreliable and even a panel of judges leaves a hard residue: when judges disagree, majority voting discards the conflict instead of resolving it. We present JuryFlow, a disagreement-guided, human-in-the-loop multi-agent evaluation framework that treats inter-judge disagreement not as noise to be averaged away, but as a precise, claim-level signal indicating where an evaluation is uncertain. JuryFlow decomposes each candidate response into atomic claims, has a panel of heterogeneous judges assign per-claim verdicts, and builds a disagreement graph whose nodes are scored by verdict entropy and whose edges encode structural similarity between claims. A human acts as a structural guide, selecting which disagreement to resolve through a single, minimal intervention rather than re-labeling the response, after which the focal claim is re-evaluated, the correction propagates along graph edges and to historically similar cases, and is crystallized into reusable rubric entries that all judges inherit, making the evaluator progressively self-refining. To enable large-scale, reproducible benchmarking without human studies, we evaluate JuryFlow in an automatic configuration in which focal selection is made by entropy ranking. On MT-Bench and LLMBar, JuryFlow improves agreement with gold labels over single-judge and majority-vote panel baselines, and ablations isolate the contributions of disagreement-targeted re-evaluation, propagation, and rubric induction. We contribute (1) a human-in-the-loop paradigm that recasts the human from labeler to structural guide, (2) the JuryFlow framework operationalizing it through a disagreement graph, focal re-evaluation, and closed-loop rubric induction, and (3) an evaluation protocol with ablations that isolate where the gains originate.
作者Viraj Parimi, Shao-Hung Chan, Han Zhang, Jingkai Chen, Brian Williams
Multi-Agent Path Finding (MAPF) aims to find collision-free paths for multiple agents in a shared environment. Classical MAPF assumes that all static obstacles are known in advance, but real-world environments can change unexpectedly due to fallen objects, spills, or other local disturbances. When such changes are spatially correlated, an observation can inform traversability estimates beyond the observed location. Prior approaches address uncertainty in traversability through contingent plans or replanning based on direct observations, but do not leverage this spatial dependence to infer the traversability of nearby unobserved locations. As a result, they cannot use one observation to anticipate nearby unobserved obstacles that may cause costly rerouting later. We focus on Belief-Aware MAPF, where map discrepancies are fixed during execution but initially unknown, and observations can be informative beyond the observed location. We propose Multi-Agent Gaussian belief Inference for Coordination (MAGIC), a framework that updates a shared belief about traversability online based on agents' observations. MAGIC uses a Gaussian Markov Random Field and Gaussian Belief Propagation to approximately infer traversability and construct detour-aware costs for standard MAPF planners. Our experiments on MAPF benchmarks show that MAGIC reduces the executed sum of costs compared to existing approaches on 96.3% of instances, across several planner families and teams of up to 800 agents, demonstrating its applicability to large-scale MAPF problems.
展开完整摘要收起摘要↓
Multi-Agent Path Finding (MAPF) aims to find collision-free paths for multiple agents in a shared environment. Classical MAPF assumes that all static obstacles are known in advance, but real-world environments can change unexpectedly due to fallen objects, spills, or other local disturbances. When such changes are spatially correlated, an observation can inform traversability estimates beyond the observed location. Prior approaches address uncertainty in traversability through contingent plans or replanning based on direct observations, but do not leverage this spatial dependence to infer the traversability of nearby unobserved locations. As a result, they cannot use one observation to anticipate nearby unobserved obstacles that may cause costly rerouting later. We focus on Belief-Aware MAPF, where map discrepancies are fixed during execution but initially unknown, and observations can be informative beyond the observed location. We propose Multi-Agent Gaussian belief Inference for Coordination (MAGIC), a framework that updates a shared belief about traversability online based on agents' observations. MAGIC uses a Gaussian Markov Random Field and Gaussian Belief Propagation to approximately infer traversability and construct detour-aware costs for standard MAPF planners. Our experiments on MAPF benchmarks show that MAGIC reduces the executed sum of costs compared to existing approaches on 96.3% of instances, across several planner families and teams of up to 800 agents, demonstrating its applicability to large-scale MAPF problems.
作者Minye Shao, Chaohui Yu, Yixuan Wu, Fan Wang, Ling Shao, Yang Long
Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark built on UK Biobank, comprising 50,401 clinical questions derived from real-world ICD-10-coded diagnoses of 4,739 participants. Each question links to a patient-specific environment containing clinical context and multi-sequence MRI from baseline and follow-up visits, where agents autonomously select which visits, organs, modalities, slices, and specialist tools to inspect and compare. Supervised fine-tuning transfers evidence-seeking workflows from 14,734 frontier-model interaction trajectories, followed by agentic reinforcement learning on the learner's own environment interactions. Privileged on-policy self-distillation and rubric-based LLM feedback refine evidence-to-conclusion reasoning without prescribing tool sequences. Experiments show that CASE moves beyond question-answer imitation toward transferable investigation policies, strengthening evidence-grounded longitudinal reasoning. Under matched evaluation conditions, our Qwen3-VL-8B based agent achieves over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8. Code will be available at https://github.com/VinyehShaw/CASE.
展开完整摘要收起摘要↓
Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark built on UK Biobank, comprising 50,401 clinical questions derived from real-world ICD-10-coded diagnoses of 4,739 participants. Each question links to a patient-specific environment containing clinical context and multi-sequence MRI from baseline and follow-up visits, where agents autonomously select which visits, organs, modalities, slices, and specialist tools to inspect and compare. Supervised fine-tuning transfers evidence-seeking workflows from 14,734 frontier-model interaction trajectories, followed by agentic reinforcement learning on the learner's own environment interactions. Privileged on-policy self-distillation and rubric-based LLM feedback refine evidence-to-conclusion reasoning without prescribing tool sequences. Experiments show that CASE moves beyond question-answer imitation toward transferable investigation policies, strengthening evidence-grounded longitudinal reasoning. Under matched evaluation conditions, our Qwen3-VL-8B based agent achieves over 16% and 10% relative improvements in answer accuracy over GPT-5.4 and Claude Opus 4.8. Code will be available at https://github.com/VinyehShaw/CASE.
作者Pranjal Aggarwal, Lawrence Keunho Jang, Sean Welleck, Daniel Fried, Ruslan Salakhutdinov, Jing Yu Koh
Computer use agents (CUAs), which use graphical user interfaces (GUIs) to complete tasks on a computer, have recently surpassed human performance on many standard benchmarks, including difficult long-horizon tasks. Their capabilities are undoubtedly impressive, however, a key barrier to the widespread adoption and deployment of CUAs remains their speed and cost. Progress towards faster yet capable CUAs requires reliable evaluation of their speed, but many CUA benchmarks currently face a reproducibility crisis. Benchmarks are based on complex infrastructure with varying machine and container configurations that confound the evaluation of the execution speed of CUAs. Towards addressing this gap, we propose cua-speedrun, which introduces standardized infrastructure and task sets, with a focus on evaluating the speed and efficiency of CUAs. cua-speedrun uses a uniform virtual machine setup and execution pipeline, along with a common agent interface that enables single-agent implementations to operate seamlessly across different benchmarks. Across four different CUA benchmarks, we evaluate how reasoning effort, agent harnesses, and environment latency affect performance, speed, and cost. We find no single model family is optimal for all three; none of the open-weight models are on the frontier, and also, unintuitively, for some models increasing the reasoning effort can speed up task completion, while faster environment input-output can slow down overall task completion time. We also demonstrate that we can effectively reduce the evaluation task set of most CUA benchmarks without degrading overall statistical power, allowing for more efficient benchmarking and comparison. We believe cua-speedrun will enable structured progress towards fast, efficient CUAs, unlocking new real-world use cases and applications. All code, infrastructure, and analysis are available at https://cuaspeedrun.com.
展开完整摘要收起摘要↓
Computer use agents (CUAs), which use graphical user interfaces (GUIs) to complete tasks on a computer, have recently surpassed human performance on many standard benchmarks, including difficult long-horizon tasks. Their capabilities are undoubtedly impressive, however, a key barrier to the widespread adoption and deployment of CUAs remains their speed and cost. Progress towards faster yet capable CUAs requires reliable evaluation of their speed, but many CUA benchmarks currently face a reproducibility crisis. Benchmarks are based on complex infrastructure with varying machine and container configurations that confound the evaluation of the execution speed of CUAs. Towards addressing this gap, we propose cua-speedrun, which introduces standardized infrastructure and task sets, with a focus on evaluating the speed and efficiency of CUAs. cua-speedrun uses a uniform virtual machine setup and execution pipeline, along with a common agent interface that enables single-agent implementations to operate seamlessly across different benchmarks. Across four different CUA benchmarks, we evaluate how reasoning effort, agent harnesses, and environment latency affect performance, speed, and cost. We find no single model family is optimal for all three; none of the open-weight models are on the frontier, and also, unintuitively, for some models increasing the reasoning effort can speed up task completion, while faster environment input-output can slow down overall task completion time. We also demonstrate that we can effectively reduce the evaluation task set of most CUA benchmarks without degrading overall statistical power, allowing for more efficient benchmarking and comparison. We believe cua-speedrun will enable structured progress towards fast, efficient CUAs, unlocking new real-world use cases and applications. All code, infrastructure, and analysis are available at https://cuaspeedrun.com.
Recent autonomous machine learning engineering (MLE) agents have made significant progress on public leaderboards. Often motivated by progress stagnation over long-horizon cycles and limited Large Language Model (LLM) primitives, modern MLE agents are deployed on top of increasingly elaborate machinery: multi-agent orchestrators, dedicated retrieval subagents, and more. While such harnesses expand, the use of more primitive but improved coding agents - where LLMs have direct access to the execution environment through read, write, and bash primitives - has received little attention in the field. In this paper we find that, under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline, pointing to the backbone as the primary driver for performance. Via a series of large-scale systematic ablation studies, we argue that the machinery layers become redundant in the coding agent setting. We conclude that the effort spent elaborating hand-crafted harnesses around strong models yields poor returns for current MLE benchmarks.
展开完整摘要收起摘要↓
Recent autonomous machine learning engineering (MLE) agents have made significant progress on public leaderboards. Often motivated by progress stagnation over long-horizon cycles and limited Large Language Model (LLM) primitives, modern MLE agents are deployed on top of increasingly elaborate machinery: multi-agent orchestrators, dedicated retrieval subagents, and more. While such harnesses expand, the use of more primitive but improved coding agents - where LLMs have direct access to the execution environment through read, write, and bash primitives - has received little attention in the field. In this paper we find that, under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline, pointing to the backbone as the primary driver for performance. Via a series of large-scale systematic ablation studies, we argue that the machinery layers become redundant in the coding agent setting. We conclude that the effort spent elaborating hand-crafted harnesses around strong models yields poor returns for current MLE benchmarks.
Auto-research agents, LLM systems that propose, implement, train, and evaluate model changes across iterations, promise to automate applied ML's experimental loop. Over long horizons, execution accuracy is a binding constraint: a change can silently leak held-out data, omit normalization, disconnect a gradient, or leave a train/eval flag unwired, invalidating expensive runs and compounding error across iterations. We present RankEvolve, an auto-research framework for evolving generative ranking models. An Executable Operating Protocol (EOP) declares phases, gates, branches, and loops, and the runtime enforces the compiled state machine. A meta-meta-harness composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one another's work. In a budget-matched evaluation, heterogeneous composition raises all-oracle execution accuracy from the best single-product baseline of 45.8 percent to 62.5 percent (paired +16.7 points, 95 percent CI [6.6, 26.7]) while achieving a 10.4 percent silent critical-defect rate. An implemented knowledge layer carries findings, including negative results, across iterations. In a twelve-iteration deployment on the open-source HSTU recommender, RankEvolve reported NDCG@10 of 0.2192 on MovieLens-20M LARGE (+4.48 percent over the published anchor) and 0.1948 on BASE (+2.80 percent). ExecML-HSTU, seeded by incidents from that deployment, provides the oracle benchmark for the execution-accuracy evaluation. A pre-specified LitGPT transfer split replicates the heterogeneous-composition effect beyond recommendation (+12.5 points, 95 percent CI [3.0, 22.0]), and a paired ablation isolates per-step from full-protocol instruction injection. These results characterize when runtime-controlled composition of coding-agent products improves execution accuracy.
展开完整摘要收起摘要↓
Auto-research agents, LLM systems that propose, implement, train, and evaluate model changes across iterations, promise to automate applied ML's experimental loop. Over long horizons, execution accuracy is a binding constraint: a change can silently leak held-out data, omit normalization, disconnect a gradient, or leave a train/eval flag unwired, invalidating expensive runs and compounding error across iterations. We present RankEvolve, an auto-research framework for evolving generative ranking models. An Executable Operating Protocol (EOP) declares phases, gates, branches, and loops, and the runtime enforces the compiled state machine. A meta-meta-harness composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one another's work. In a budget-matched evaluation, heterogeneous composition raises all-oracle execution accuracy from the best single-product baseline of 45.8 percent to 62.5 percent (paired +16.7 points, 95 percent CI [6.6, 26.7]) while achieving a 10.4 percent silent critical-defect rate. An implemented knowledge layer carries findings, including negative results, across iterations. In a twelve-iteration deployment on the open-source HSTU recommender, RankEvolve reported NDCG@10 of 0.2192 on MovieLens-20M LARGE (+4.48 percent over the published anchor) and 0.1948 on BASE (+2.80 percent). ExecML-HSTU, seeded by incidents from that deployment, provides the oracle benchmark for the execution-accuracy evaluation. A pre-specified LitGPT transfer split replicates the heterogeneous-composition effect beyond recommendation (+12.5 points, 95 percent CI [3.0, 22.0]), and a paired ablation isolates per-step from full-protocol instruction injection. These results characterize when runtime-controlled composition of coding-agent products improves execution accuracy.
Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. Because repeatedly executing long-horizon computer-use tasks is costly and stochastic, we further propose an execution-free, task-conditioned evaluation framework that projects the original task grader onto prompt-addressable intentions and measures how much task-relevant information is retained after the voice interface. On WildClawBench, Talk2Agent's execution-free native projection provides a practical, execution-grounded measure of voice-interface quality, correlating with downstream task completion and improving Pearson correlation by 0.246 over WER/CER on 32 hours of real human speech.
展开完整摘要收起摘要↓
Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. Because repeatedly executing long-horizon computer-use tasks is costly and stochastic, we further propose an execution-free, task-conditioned evaluation framework that projects the original task grader onto prompt-addressable intentions and measures how much task-relevant information is retained after the voice interface. On WildClawBench, Talk2Agent's execution-free native projection provides a practical, execution-grounded measure of voice-interface quality, correlating with downstream task completion and improving Pearson correlation by 0.246 over WER/CER on 32 hours of real human speech.