Agent 进展

Urban navigation requires embodied agents to pursue long-horizon goals through local decisions based on egocentric observations. However, existing agentic navigation methods often struggle to translate distant goals into coherent local decisions in large-scale physical environments. Their reliance on linguistic reasoning over transient observations or limited history constrains anticipation of the consequences of actions and future conditions, despite the importance of such foresight for navigating long and complex urban routes. To bridge this gap, we propose PreAct-Nav, an agentic navigation framework that equips frozen policies with anticipatory reasoning for robust urban navigation. Our central idea is to anchor local decisions in persistent medium-horizon subgoals, assess the consequences of predicted actions before execution, and continually update the reasoning context using actual outcomes. At its core, a navigation memory module maintains the active subgoal and relevant experience across decisions, translating distant goals into actionable intermediate objectives. We further introduce a predictive world sandbox that uses an action-conditioned world model (AC-WM) to forecast world dynamics conditioned on candidate movements. A vision-language model (VLM) reasoner interprets these predictions under the current subgoal to retain or revise actions. After execution, real observations are used to assess outcomes, correct inconsistent assumptions, and update memory to continue or reformulate the subgoal. Extensive evaluations demonstrate that the proposed PreAct-Nav improves action selection through memory updates and visual prediction, with more pronounced gains on longer routes and routes with more turns.

AI agents based on foundation models (FMs) have demonstrated strong capabilities to perform complex open-ended tasks. However, they face some common challenges in practice: (a) agent behavior can deviate drastically even for semantically similar tasks, leading to catastrophically propagated errors; (b) high cost and latency due to FM calls, repeated in full whenever a task recurs with different inputs; (c) FMs' limited context and instruction following capability confine how well agents manage the ever-growing execution context and follow complex plans. We introduce $A$gent $B$ehavior as $C$ode $Agent$ (ABCAgent), which uses a symbolic program (e.g., Python code with potential neural functions) to fully specify the agent's behavior at runtime, with a powerful FM agent editing that program for flexibility. Behavior is thus specified without premature variable binding, and its execution is deterministic. We evaluate ABCAgent on six agent benchmarks, two of which we construct to test how well a derived program generalizes to variants of the task it was written for. ABCAgent matches a model-matched neural agent on GAIA and augmented GAIA, and surpasses it where robustness and long control flows matter: 98.3% against 97.3% on GSM-Symbolic ($p = 0.001$), 71.9% against 47.4% $\mathrm{Pass}^4$ on the telecom domain of $τ^2$-bench ($p = 0.0001$), and more records written correctly at every loop length on our control-flow-augmented WorkArena benchmark. For more parametric task families, ABCAgent is also significantly superior in efficiency. Without authoring a new program, ABCAgent solves 92.6% of GSM-Symbolic instances and 20.1% of augmented GAIA variants, which yields $5.2\times$ lower latency and $7.0\times$ lower cost on GSM-Symbolic, 19% lower cost on augmented GAIA, and $9.5\times$ lower agent latency on $τ^2$-telecom.

Agent Skills, the SKILL.md files that tell an LLM coding agent how a project works, are revised like code, yet what a revision does to the agent is unknown. From 2,608 first/last revision pairs of 3,159 Skills, we characterize how Skills evolve and how they change together with the configuration of the agent's harness. We then focus on rule changes, revisions that add or remove a rule we can check automatically, such as "run allium check". We measure their effect on 21 models in single answers and on four agents in a sandbox, and their cost on 20 of these models and the four agents. Most revisions (55%) change a rule or procedure, and commits that revise a Skill change harness files such as CLAUDE.md more often than other commits of the same size. Across 16 open-weight models, an added rule raises compliance in a single answer by +0.41 on average. Across the four agents, the rate at which the agent takes the required action rises by +0.23 on average (+0.16 to +0.36), and for the three agents that blind judges assessed, final correctness rises by +0.10 on average (+0.06 to +0.14). The gain comes mainly from rules that name a command or path the old Skill did not mention. Real tools load a Skill's body only when the agent decides it needs it. In that setting the four agents keep about half of the action gain on average (51%), and the three open models about 38%. A revision adds 18-19% input tokens to a single answer and no detectable cost to an agent episode, while loading a Skill's body raises the tokens of an episode by 50% on average.

Interactive language-model agents increasingly solve complex tasks through long-horizon, multi-call reasoning, where errors in beliefs or actions can compound across tool interactions. Retrospective search can recover from such failures but is prone to misallocation. Delayed outcomes obscure the contribution of intermediate search decisions, leading to Attribution Complexity, while evolving execution evidence leads to Adaptation Complexity, where previously learned estimates become stale. To address these challenges, we first introduce Search Value Dynamics (SVD), which characterizes the evolving trade-off between the gain and cost of retrospective search. Building on SVD, we propose ForkPilot, a self-evolving two-stage policy-learning framework. In the first stage, ForkPilot learns a search-value policy offline from completed trajectories through automatically constructed outcome comparisons. In the second stage, it makes search decisions based on current observations and then self-evolves by incorporating newly completed trajectories into subsequent policy updates. We evaluate ForkPilot across 6 diverse benchmarks and 7 widely used LLM backbone families, including four open-source families, GPT-5.6 Sol, and Opus 4.8 in a production agentic system, against 9 competitive baselines, including a real-world harness deployment used by hundreds of thousands of paid users. ForkPilot achieves comparable state-of-the-art performance while reducing token usage by up to 59.2%, demonstrating its efficacy.

Large language model (LLM) agents solve complex tasks through multi-step interactions with external tools. These interactions often contain recurring local tool sequences. Treating such sequences as composite "Skills" can shorten tool-use trajectories and reduce repeated low-level decisions. However, when atomic tools and composite skills coexist, skill use becomes a policy problem: the agent must decide whether the current state requires atomic fine control or skill-level abstraction. In this paper, we argue that effective skill use should be studied as adaptive tool granularity selection. The most direct training signal for this problem is to compare the consequences of atomic and skill choices available from the same state. Based on this view, we propose CIPO, a Counterfactual Imagination Policy Optimization framework for adaptive tool granularity. CIPO constructs executable skills through budget-constrained mining of successful tool-use trajectories and trains granularity decisions with counterfactual branch rollouts. For each base rollout, CIPO branches at the first eligible granularity decision and replaces the chosen action with a feasible atomic or skill alternative. The paired outcome difference serves as a supplementary reward for policy optimization. Experiments across multiple benchmarks and model backbones show that CIPO improves task success and decision efficiency over baselines. Further analyses show that CIPO learns effective skill use by improving the choice between atomic tools and composite skills based on the current state, without simply increasing skill frequency.

An agent harness is the code that organizes context, maintains state, and coordinates tool calls for a language model. We study how to improve the harness under a limited evaluation budget while keeping model weights fixed. Our method, MESH-Harness, organizes each harness into functional modules with explicit role-specific interfaces, allowing alternative implementations of each module to be substituted and recombined. It uses shared module representations and full-covariance LinUCB to score candidate combinations based on predicted performance and exploration value. Mixed-start coordinate ascent selects complete configurations for evaluation without enumerating the combinatorial space. Validation traces then guide local code edits, and the resulting candidates are incorporated into fixed-capacity role-specific pools for subsequent recombination. On text tasks, retrieval-augmented mathematical reasoning, code generation, and interactive scientific tasks, MESH-Harness outperforms Meta-Harness by 5.70, 7.01, 2.00, and 5.00 points, respectively, under matched candidate-evaluation budgets. Iterative harness optimization improves MESH-Harness by 5.63-7.79 points over its first-round configurations. For the reported configurations, aggregate test-time cost is 44.2% lower than that of Meta-Harness, while total cost including search is 14.6% lower. These results show that combining module-level design reuse with feedback-driven compositional search can systematically improve agent harnesses while keeping overall optimization cost under control.

As Large Language Models (LLMs) gain tool use capabilities such as web search, they can retrieve and cross-reference public information, creating privacy risks beyond memorization. One manifestation is re-identification: linking an anonymized interview transcript to a named individual. Yet without ground-truth identities, the coverage of such attacks and the protection offered by a defense cannot be reliably measured. We introduce AgentDOXX, an evaluation suite of 822 synthetic interview transcripts grounded in public information about real individuals with known identities. We evaluate fifteen configurations of open-weight and proprietary models, isolating the effect of web search, and analyze their search trajectories to distinguish retrieval-driven from parametric identifications. Ground-truth identities reveal that re-identification risk is distributed across an agent's execution: retrieval and parametric recall both contribute, with open-weight models identifying 15-28% of transcripts without search; identification succeeds in over 88% of cases once the target appears in a retrieved result; entity masking leaves at least one attacker successful on 85.3% of a stratified sample; and privacy instructions suppress naming but not retrieval, with configurations scoring 0% accuracy yet retrieving the subject in up to 62% of transcripts. We further show that observed attack trajectories can provide supervision for localizing identifying spans, offering a path toward attack-informed anonymization.

Long-horizon agents consume external content, invoke tools, and modify persistent state. Indirect prompt injection can exploit task-specific context, propagate across causally connected stages, and alter a consequential action while the workflow continues; we term this staged prompt injection. We build an automated, feedback-guided attack generation pipeline and apply it to Claude Code and Codex in their native runtimes. The confirmed attacks span eight workflow scenarios, seven attack goals, and six injection surfaces, showing that production agents are vulnerable to context-aware, multi-step injection over long horizons. Stopping such attacks requires a decision before each consequential action: input screening and completed-run evaluation cannot locate the intervention point, and existing pre-action methods use incompatible units and labels. We therefore formulate boundary action auditing: given initial context, a trajectory prefix, and a fully specified pending message or tool call, an auditor predicts Pass or Block before its effect occurs. Pairing attacked and benign executions yields a 479-pair, 3,112-unit benchmark. We further propose Path-Aligned Attribution (PAA), a training-free auditor that decomposes pending actions into operative elements and traces what supplied each value and guided each decision. PAA blocks only when the model attributes an unwarranted, material effect on an element to an attacker-reachable source that either provides unqualified steering or conflicts with visible evidence. Under full-benchmark fail-open scoring with Claude Sonnet 5, PAA reaches 86% Block recall at a 6-8% false-block rate (FBR), whereas ARGUS reaches 44-47% recall at 16-33% FBR. Under the same backend, on the tool calls that all three auditors natively support, PAA has higher recall and lower FBR than VIGIL and ARGUS; all paired 95% confidence intervals exclude zero.

The increasing complexity and scale of modern cyber-attacks demand intelligent and computationally efficient Intrusion Detection Systems (IDS). However, designing effective data pre-processing pipelines traditionally involves substantial trial-and-error effort and repeated evaluation of alternative configurations. For large, high-dimensional network traffic data, this process can create a significant computational burden. In this work, we propose AutoDP-LLM, an automated pre-processing framework designed to reduce manual pipeline development and computational overhead. Specifically, AutoDP-LLM leverages Large Language Models (LLMs) to autonomously generate and validate executable data pre-processing pipelines. The framework combines deterministic host-side planning with LLM-based specialist agents to formulate data-processing strategies, synthesize executable code, and adaptively determine retained feature sets using semantic reasoning and training-derived statistical evidence, without requiring a predefined feature budget. Focusing on multiclass intrusion detection, we evaluate AutoDP-LLM on the UNSW-NB15 and NSL-KDD benchmark datasets using multiple downstream classifiers. Comparative experiments against conventional feature-selection methods show that AutoDP-LLM achieves competitive detection performance while automating the generation of compact and executable pre-processing pipelines. Component-level ablation experiments further demonstrate the complementary contributions of the semantic and statistical feature-reduction components. The repeated generation, validation, execution, and assessment of candidate pipelines are amenable to parallel execution, highlighting the potential of scalable computing environments, including high-performance computing (HPC) systems, to support automated IDS pipeline development.

Small language models (SLMs) offer a promising foundation for on-device agents through low-latency, resource-efficient inference, yet limited reasoning and planning capabilities constrain their performance on long-horizon tasks requiring multi-step interaction with the environment. Step-level collaboration between SLMs and larger cloud-hosted models can bridge this gap, but identifying states that warrant cloud assistance remains challenging: the contribution of each cloud call is entangled with subsequent actions and can be assessed only from the final task outcome. Compounding this challenge, the SLM must balance two competing objectives: maximizing task success and minimizing cloud calls. To address this, we propose Sibyl, an algorithm that trains SLM agents to selectively consult cloud models at the step level and internalize their guidance for subsequent decisions, achieving strong task performance with minimal cloud reliance. Sibyl follows a three-stage training pipeline that (1) builds a robust base policy through consultation-free self-evolving reinforcement learning (RL); (2) cold-starts consultation behavior via decisive-disagreement state mining; and (3) jointly optimizes consultation decisions and guidance internalization through consultation-aware RL. Experiments on ALFWorld and WebShop demonstrate that Sibyl, using only a 0.6B-parameter model, outperforms state-of-the-art baselines, including agent training and routing methods, by 95.2% and 80.4% in success rate while averaging only 0.8 and 3.9 cloud calls per trajectory, respectively.

Human health is a continuous, dynamic trajectory shaped by the cumulative interplay of biological processes, clinical events, behaviors and environmental exposures across the life course. Unifying the full breadth of lifelong health information, including longitudinal records, genetic variation, molecular profiles and environmental histories, is essential for whole-person modeling and remains a major challenge. We introduce LifeLong Digital Twin, a unified, event-driven modeling paradigm that organizes Life Events into daily Health States and accumulates them into Lifelong Health Context. An accompanying Agent Harness incorporates multimodal evidence beyond the language model's textual context. We evaluate four language models across 25 disease endpoints on three tasks: Disease Trajectory Forecasting, Disease Risk Ranking and Multi-horizon Disease Prediction. The approach yields marked gains over the reference condition: model-averaged F1 increases by 22.0% for disease identification in trajectory forecasting and 18.3% for five-year disease outcomes; thyroid-disease F1 reaches 0.669. The framework provides a foundation for whole-person digital twins and research on personalized lifelong disease prevention.

LLM agents for coding, search, and workplace tasks increasingly rely on long-context capabilities to effectively aggregate and reason over extended interaction histories. Recent work has incorporated agent trajectories into mid-training stage, drawing on their naturally long and interaction-rich structure. Yet how to organize the information within these trajectories into effective mid-training supervision remains underexplored. In this work, we investigate the relationship between long-context and agent atomic capabilities and introduce TrajLong, a novel framework that compiles trajectories into long-context training tasks with dense supervision, targeting three representative atomic capabilities: evidence grounding, cross-evidence aggregation, and temporal state maintenance. We mid-train Qwen3-14B-Base and Qwen3-30B-A3B-Base with data compiled by TrajLong, followed by supervised fine-tuning. Experiments on 6 long-context and 12 agent benchmarks demonstrate broad performance gains, with controlled ablations showing improvements over raw and masked trajectory baselines. Capability-level analyses further reveal task-dependent associations between long-context and agent atomic capabilities. These findings suggest that the shared capability demands of long-context reasoning and agent execution provide a principled basis for designing mid-training data to develop downstream agent capabilities.

LLM-based agentic systems are emerging as a new software paradigm. Modern agents are typically composed of backbone LLMs and a surrounding harness that serves as the operational software infrastructure for agent execution. As agent harnesses grow increasingly complex, agents suffer from diverse harness implementation bugs, raising substantial reliability concerns. In this work, we conduct the first empirical study to systematically investigate the test adequacy of harness in real-world agentic systems. Our analysis reveals that agent harness remains substantially undertested. In particular, LLM-dependent harness (LDH) code, despite its critical role in processing LLM outputs and governing agent behavior, receives limited testing attention, with less than half of its lines and branches covered by existing tests. Motivated by these findings, we further propose HarnessTester, the first harness-oriented test generation technique that incorporates explicit agent-harness contract support to construct contract-faithful test setups and extensively exercise LDH code. Our evaluation shows that HarnessTester substantially outperforms state-of-the-art general-purpose test generation techniques in achieving 75.95%/84.76% larger line/branch coverage gains and 69.89% larger mutation-score gains. Furthermore, HarnessTester detects 122 real-world harness bugs in widely-used agentic systems (e.g., OpenClaw), among which, 88 bugs are previously-unknown bugs and 69 bugs have been confirmed by agent developers. These results highlight the practical effectiveness of HarnessTester in improving test adequacy and assuring the reliability of real-world agentic systems.

Large language model (LLM) agents are increasingly explored for root cause analysis (RCA) in microservice systems, yet empirical guidance on how to design and combine their tools remains limited. We conduct a controlled empirical study of tool abstraction and composition across models and microservice environments. We implement 24 structured tools for metric access (L1), evidence analysis (L2), and diagnosis (L3), alongside a Python-based reference setting (L0). Using 375 failure cases from three microservice systems, we evaluate eight configurations with Qwen3.7-Plus and compare four Qwen models on a shared subset of four configurations. Our results show that tool configurations affect root cause localization and failure type identification differently. With Qwen3.7-Plus, L3 achieves 82.8% top-1 localization accuracy compared with 85.4% for L0, while requiring less than half the time per case. Adding tool levels can improve type identification while reducing localization accuracy. Trajectory analysis reveals cases in which agents override correct diagnostic recommendations after consulting additional evidence. Model choice also changes tool benefits: adding L3 to L1+L2 improves diagnosis for three models but hurts Qwen3-8B, which rarely invokes L3. Tool configuration rankings further change across microservice systems. These findings provide an empirical foundation for designing and using RCA tools, guiding tool selection and composition according to the model, diagnostic objective, and target system.

Structured radiology reporting mitigates the heterogeneity of free-text reports, yet its benefits depend on high-quality reporting templates. In practice, such templates are conventionally built through labor-intensive expert consensus and therefore vary across institutions and lag behind evolving clinical practice. Large language models (LLMs) enable automated template induction, but existing approaches remain limited: single-LLM induction is constrained by context length, and the corpus-scale method ASTAR produces a static, closed-corpus template without external grounding or downstream adaptation. To address these limitations, we propose TEMPLAR, a TEMPLate-centric Agentic framework for inducing and evolving standardized Radiology reporting templates from large-scale clinical corpora. TEMPLAR treats the template as a persistent central state maintained alongside two provenance-aware knowledge graphs, namely an anatomical graph that constrains template construction and a diagnostic graph that supports finding-to-diagnosis reasoning. Three agents operate on this state. The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms via dual-view similarity clustering; the Evolution Agent then assembles these slots into a hierarchical template and revises it under consistency constraints, external clinical evidence, and downstream structuring feedback; and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnostic reasoning. Across four datasets, TEMPLAR outperforms ASTAR, three medical LLMs, and six general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity, while achieving the highest or tied-highest LLM-rated template quality. Its fidelity advantages over ASTAR persist under cross-dataset transfer, and cumulative ablations support complementary contributions of its key components.

Small language model (SLM) agents need safety controls that track consequences across tool calls with little monitoring overhead. A private read, for example, becomes a leak when a later action sends that data outside the system. We introduce CARB (Conformal Agent Risk Budget), which calibrates when to stop an agent using a ledger of harm incurred before stopping. Under exchangeable episodes, standard conformal risk control bounds this declared loss in expectation over calibration and a future episode. G-CARB selects scorer evidence along observable dependencies from private sources to outgoing actions. The ledger still covers the entire executed history, and computing the gate score requires no additional language-model inference. On AgentDojo replay with two 14B backbones, G-CARB roughly halves scorer-input records at intermediate risk budgets while improving autonomous task completion relative to full-prefix scoring; random context of the same size achieves similar gains. Controlled examples show how retaining the relevant dependency can further avoid stopping benign work.

When an agentic prover works on an open problem, there is no proof assistant to fall back on: its verifier and lemma library are ultimately language models judging model outputs. We instrumented such a system end to end and analyzed $51{,}754$ traced observations across three full runs ($186$ hours, \$$5{,}694$). We find three connected failure modes. First, the three-model verifier requires unanimity and treats parse or API failure as non-approval; in $10$ of $12$ verification events, one member returned no parseable output or an API error, making acceptance arithmetically impossible without surfacing an error. Second, when the ensemble did function, one verifier approved $3$ attempts that GPT rejected, each claiming to resolve the open problem; a single-verifier design would therefore have announced a solution three times. Third, because nothing could be approved, every review was a refutation, yet the lemma extractor mines reviews as well as proofs: $24$ of $93$ lemmas ($26%$) were extracted from rejected arguments with their refutational context removed. Taken together, these findings show that without external verification, supervision is itself a critical trust boundary: systems must distinguish abstention from rejection, preserve useful disagreement, and preserve the provenance and polarity of information before it becomes future context.

Smart homes and smart devices are becoming more prevalent across millions of homes around the world. With the rise of AI, the smart home industry is quickly increasing its integration to manage common smart home tasks. However, existing work in large language models (LLMs) as agents within smart homes have shown minimal resilience due to limited environment scenarios or poor performance in complex tasks. We explore several strategies for LLMs as agents within the popular open-source software (OSS) smart home automation framework HomeAssistant to increase overall smart home resilience. Our approach combines traditional supervised learning techniques and optimized prompting as the core learning process. We use a diverse set of LLMs covering different levels of reasoning and costs in our optimization pipeline and evaluate their performance on a subset of the hardest tasks in a smart home benchmark. We include ReAct and Reflexion agentic paradigms and reveal how both provide marginal return on investment compared to fine-tuned LLMs for multi-device control within HomeAssistant but show promise in resilience tasks such as failure response.

Infrared small-target detection plays an important role in maritime monitoring and aerial surveillance. Although multimodal large language models (MLLMs) offer promising capabilities for visual understanding, existing MLLM-based approaches struggle to precisely localize infrared small targets. In this paper, we propose IRSTD-Agent, an agentic framework for infrared small target detection through dynamic visual search. The framework enables an MLLM to adaptively determine where and at what scale to inspect an image and progressively gather fine-grained visual evidence for precise target localization. Five complementary visual tools (PROPOSAL, ZOOM, DETECT, DROP and REFINE) support object candidate discovery, adaptive observation, target localization, hypothesis rejection, and target extent refinement, together enabling a coordinated search process over original-resolution images. To teach the MLLMs to conduct this search, we introduce Zoom-guided Interaction Learning, which uses annotation-derived interaction trajectories to supervise tool selection and the corresponding arguments. Through extensive experiments on WideIRSTD-Full and IRSTD-1k datasets, we demonstrate that IRSTD-Agent outperforms the evaluated vision-language models and enhances the precise localization capabilities of MLLMs in IRSTD tasks.

Computer-use agents need to capture procedural knowledge of how people use software. User telemetry offers a scalable source of this knowledge. However, learning reusable skills from these logs requires addressing three challenges: (1) Goal Underspecification, since logs do not record the goal behind each action; (2) Non-Replayability, since past activity cannot be replayed to evaluate skill updates; and (3) Interleaved Trajectories, since logs may mix several tasks without marking their boundaries. To address these, we introduce TeleTune, a framework for learning a textual skill library from offline logs without recorded goals, cannot be replayed during optimization, and may interleave tasks. TeleTune uses action-prediction errors on logged trajectories to propose library edits and keep only those that improve held-out action-prediction accuracy, which we call skill-guided progress. The learned workflows also enable retrieval of demonstrations that cover the subgoals of a new task. At test time, the agent is provided with the learned library and the workflow-based retrieved demonstrations. Experiments on WorkArena and Online-Mind2Web show that TeleTune outperforms random retrieval, Agent Workflow Memory (AWM), and their combination. We find that the best baseline varies by setting, whereas TeleTune achieves average success rates of 77.1% and 80.6%, respectively, improving over the strongest baseline on each benchmark by 6.7% and 7.7%. Under the heaviest perturbation of the WorkArena training data,TeleTune keeps the highest average success rate at 68.5%, 6.3% above the strongest baseline. Our analyses show (1) skill optimization and workflow-based retrieval are complementary, (2) optimizing on fixed logs costs 5 to 75 times fewer tokens than validating the same edits with live episodes, (3) skill-guided progress tracks the live success rate.

Proof auto-formalization translates natural-language (NL) theorems and proofs into a formal language (FL) such as Lean, enabling mechanical verification. Despite rapid progress, research-level proofs often depend on concepts missing from leading proof assistant libraries (e.g., Lean's Mathlib), and successful compilation does not guarantee that a translation preserves the theorem's meaning or the proof's reasoning. Furthermore, aligned NL-FL training data are scarce, and leading agents often rely on costly frontier models and manually engineered harnesses. To address the above issues, we present AIProver, an agentic framework for autonomous proof auto-formalization and proof synthesis (AFPS) that jointly post-trains a 119B open-weight language model and evolves its agentic, tool-calling harness with HarnessEvolve. Verifiers assess type correctness, proof completeness, and semantic correctness, returning rewards and diagnostic certificates that drive model fine-tuning and alternating reinforcement learning via symbolic feedback and HarnessEvolve, a certificate-driven evolutionary search over the whole harness control flow that re-tailors the harness to the updated model. For research-level training and evaluation, we introduce LoCoBench, 58.9k instances from Mathlib, CSLib, Mizar Math Library, and a bounded-arithmetic textbook, with a 771-instance validation split whose theorem-proof pairs have no public Lean formalization. Against 39 frameworks spanning AFPS agents, frontier LLMs, and coding agents, AIProver lifts pass@4 semantic correctness over its Leanstral-1.5 base from 15.7% to 36.7% and outperforms every other open-weight system and Aristotle. As a Claude Code and Codex skill, it lifts their semantic correctness from 41.9% and 34.1% to 79.8% and 62.4%, respectively. Further, it is also 24% cheaper than Numina-Lean-Agent, pushing the accuracy-cost frontier of research-level AFPS.

Generating a high-quality gaming world from a natural-language game intent requires joint reasoning about scene structure, spatial layout, gameplay objectives, interactive entities, and executable gameplay logic. Existing coding agents can generate individual assets, scenes, or scripts, but often struggle to maintain consistency across these components. We propose Code2Games, an agentic framework that builds a structured gaming world upon a base Blender world generated from the same game intent. Code2Games coordinates scene analysis, gameplay planning, constrained gaming-world generation, and gaming-engine customization through a shared scene-gameplay representation with persistent element correspondence. After world generation, Code2Games adapts the generated world to Unreal Engine 5 and employs an execution-guided reconstruction process that uses compilation diagnostics, runtime feedback, and gameplay test results to resolve inconsistencies arising during engine adaptation. To systematically evaluate gaming-world generation, we introduce the GameCode4D benchmark, which comprises ten fixed game prompts spanning different levels of scene and gameplay complexity. We evaluate the generated results across four dimensions: visual quality, interactive fidelity, multimodal artifact quality, and playable-game quality. Experiments demonstrate that, compared with direct gaming-world generation by coding agents and existing baseline methods, Code2Games consistently improves the visual quality and interactive fidelity of generated gaming worlds, as well as the quality of the resulting games after engine adaptation.

Search agents repeatedly make short decisions about relevance, evidence sufficiency, and search actions. Using generative language models for these decisions introduces latency and unreliable confidence. We present SearchJev, a fast and calibrated System-1 model that separates search decisions from System-2 reasoning and generation. Given a search state and a decision schema, SearchJev directly scores legal options without autoregressive output generation. We propose Soft-Label Learning for Calibrated Decisions (SLCD) to learn decision probabilities from uncertain supervision and calibrate their confidence. In a dual-system search agent, SearchJev handles short decisions and delegates uncertain judgments to System 2, which retains planning, query generation, and answer composition. We also introduce SearchDecision-Bench, a benchmark unifying six types of search decisions for training and evaluation. On SearchDecision-Bench, SEARCHJEV improves decision quality over same-size Qwen3.5 autoregressive models, achieves 5.2-5.3 times faster decisions, and reduces average expected calibration error by 41-74%. On BrowseComp-Plus, the dual-system agents achieve a 3.7-4.7 times speedup in active search time while improving answer accuracy from 45% to up to 54%.

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

Autonomous research seeks sustained model improvements through iterative experimentation and feedback. LLM agents show promise in automating machine learning and language-model post-training, but their ability to sustain multimodal improvement remains unclear. We introduce MMPostTrainBench, a benchmark spanning eight tasks in image, audio, video, and joint audio-video understanding and image-grounded software repair. Agents operate from a common base model within fixed budgets, using development feedback before independent evaluation of their submitted models. Evaluation covers target and non-target model outcomes, iterative model improvement and selection, and research integrity. Across all eight tasks, 52.1% of model--task means fall below the base, and evaluated submissions also exhibit non-target regressions. Model performance does not consistently improve across research iterations, and agents do not reliably select the best evaluated candidate for submission; final submissions trail that candidate by up to 5.38 percentage points. Extending autonomous research from text-only to multimodal tasks introduces additional sources of error in perception, cross-modal alignment, and temporal grounding. The observed regressions and selection gaps highlight the need to balance targeted improvements with non-target capability preservation and to retain gains across research iterations. These requirements motivate MMResearch, a multimodal research framework that connects media-grounded evidence to hypotheses and interventions, carries findings across rounds through hierarchical memory, and retains candidates using development evaluation. Added to existing code-agent runtimes, it improves submitted-model accuracy by up to 7.75 percentage points for Claude Opus 4.8 with Claude Code and 2.33 points for GPT-5.6-sol with Codex.

Natural language to SQL (NL-to-SQL) benchmarks are foundational to progress in data analysis research, yet recent work has shown that widely-used benchmarks contain significant annotation errors. These errors silently corrupt evaluation metrics, penalize correct model output, and distort the field's understanding of state-of-the-art performance. We present SALUS, a system that automatically detects annotation errors in NL-to-SQL benchmarks. SALUS frames benchmark auditing as a weakly supervised error detection: SQL generated by multiple LLM agents drive a suite of complementary weak-labeling functions. By passing this noisy vote matrix through a generative label model, we extract high-confidence training samples without requiring human ground truth. These samples train a decision plane that maps gold SQL query features to per-agent trustworthiness, allowing SALUS to intelligently fuse reliability estimates with raw verdicts for rigorous benchmark error detection. We evaluate on BIRD-Clean-xs, a benchmark of 298 BIRD development tasks with manually verified correctness labels. SALUS achieves F1 = 0.9194, significantly outperforming the state-of-the-art baselines. Applying SALUS to the full development sets, we estimate annotation error rates of approximately 37% on BIRD and 27% on Spider.

LLM agents combine reasoning, tool use, and persistent memory to support work across tasks by reusing stored operational records as premises for later actions. However, environmental or requirement changes can invalidate these records, while existing action review, provenance tracking, and clarification mechanisms may leave the underlying persistent state uncorrected. Our audit of coding-agent trajectories identifies candidate failure chains in which invalid records are reused, leading to task failures and unsafe modifications. We propose StateWise, a framework for diagnosing and repairing persistent operational state before action execution. StateWise uses record-level counterfactual replanning to identify decision-critical records, then establishes their current validity through reliability checks, read-only verification of machine-observable facts, and targeted clarification of developer-owned intent. Typed evidence grounding binds evidence to specific records and scopes, enabling persistent corrections with repair lineage. The agent then replans from the repaired state, followed by an independent state-action check before execution. We evaluate StateWise on 150 executable coding-agent cases across diverse runtime environments, workspace configurations, and repository settings, complemented by cross-model evaluations. Under corrupted persistent state, StateWise achieves 93.3% overall correctness, compared with 38.7% for the baseline agent, with no unsafe actions. Component ablations, multi-task experiments, and transfer evaluations further demonstrate effective recovery, persistent corrections, and transferability across repositories and tool interfaces.

Large language models for code generation often fail on execution, multilingual coverage, and contamination control, especially under frozen backbone constraints. We present CodeForge-MA, a unified framework that improves code synthesis through a multi-agent data forge, execution verified reinforced instruction tuning, and a language conditioned mixture of LoRA adapters. Four specialized agents, Composer, Reviewer, Executor, and Curator, iteratively refine instruction code pairs, validate them with tests, and filter duplicates and benchmark leakage. During training, we combine masked supervised fine tuning with a test driven reinforcement objective to align generations with executable correctness. For the larger model, we use sparse expert routing over low rank adapters to improve cross language transfer while keeping the base model unchanged at inference. Experiments show that joint data, objective, and adapter design yields robust gains across programming languages.

Multi-agent debate (MAD) has emerged as an effective paradigm to improve the reasoning capabilities of large language models (LLMs) and is increasingly being extended to multimodal settings. However, existing multimodal MAD frameworks typically expose agents to the same fixed visual input, ignoring substantial variation in the visual scale needed across samples and agents. In addition, these frameworks frequently suffer from groupthink, a phenomenon where agents prematurely abandon correct deductions to conform with confident but hallucinated peer responses. To address these bottlenecks, we introduce DREAM (Dynamic Resolution Assignment For Multimodal Multi-Agent Debate), which operates via two core components: (1) Dynamic Resolution Assignment, a zero-shot probe round where agents test multiple resolutions, quantify uncertainty using Average Normalized Log-Likelihood (ANLL), and use an adaptive threshold to assign each agent to its empirically optimal resolution; (2) Uncertainty-Guided Rollback Aggregation counters groupthink by tracking each agent's uncertainty over rounds and restoring early low-uncertainty answers overridden by group pressure. On six multimodal datasets, DREAM improves the accuracy-token trade-off over multi-agent debate baselines by 1.5-3.2% accuracy without dataset-specific tuning.

Large language model (LLM)-based multi-agent systems increasingly rely on memory to transform execution trajectories into reusable procedural knowledge. Yet repeated retrieval also makes memory errors persistent: memory pollution arises when outdated, weakly supported, or spuriously successful procedures become recurring components of future reasoning. Multi-agent execution introduces an additional structural risk. Scope collapse occurs when procedural knowledge escapes the coordination scope in which it was shown effective and is repeatedly reused at incompatible decision levels, allowing local errors to influence cascades of downstream decisions. Meanwhile, task-level failures provide ambiguous supervision because they rarely reveal which recalled knowledge was responsible. We introduce AECG, a framework for asymmetric experience consolidation and governance for multi-agent systems. AECG turns memory from static experience storage into a dynamic reliability-governance loop, preserving coordination scope and using multi-scale, confidence-aware reliability to detect degradation. It then combines degradation with downstream impact to prioritize high-risk knowledge under a bounded review budget, applies targeted interventions, and reactivates revised skills only after paired replay. Across three multi-agent frameworks and four benchmarks, AECG achieves the best score in 11 of 12 framework--benchmark settings and improves over the strongest competing memory method by as much as 10.23 percentage points; removing scope preservation reduces accuracy by up to 16.89 points. AECG thereby reframes multi-agent memory from passive accumulation into auditable reliability governance. Code is available at https://github.com/fenhg297/AECG