Agent 进展

Large language models (LLMs) can influence people's beliefs, yet little is known about whether and how they can manipulate each other. To investigate this, we simulate conversations between two agents: a target LLM that role-plays a human persona based on demographic and psychological attributes, and an influencer LLM that aims to make the target's beliefs more extreme. We examine radicalization along two pathways: resonance, where the influencer reinforces a target's pre-existing belief, and persuasion, where the influencer promotes a belief the target initially considers unimportant. Across affective and behavioral metrics, we find that both mechanisms radicalize the target. However, resonance produces consistently stronger effects than persuasion. Different influence tactics, such as using sycophancy and unverified claims, produce different levels of radicalization, but not consistently across metrics. We further show that resonance propagates to related beliefs, suggesting interconnected belief structures within AI agents. These findings indicate that AI agents are susceptible to radicalization, particularly when messages align with their existing beliefs, raising concerns about the vulnerability of personalized AI agents and multi-agent AI ecosystems.

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

Advancing beyond traditional static scoring models, LLM-powered agentic recommender systems (LLM-ARS) instantiate users and items as autonomous agents, whose semantic states are dynamically refined through a recurrent process known as collaborative reflection. While this mechanism improves recommendation quality, it simultaneously introduces a systemic vulnerability: adversarial evidence injected into a single agent can be rationalised into a legitimate preference narrative, written back into memory, and propagated to other agents through interaction contexts. We term the local rationalisation process reflection laundering, and its system-wide escalation through collaborative reflection collaborative-reflection hijacking. Existing attacks on recommender systems, whether based on interaction-level data poisoning or text-level adversarial perturbations, assume static pipelines and thus cannot exploit this recurrent, multi-agent amplification pathway. To bridge this gap, we first conduct a controlled vulnerability analysis that establishes two exploitable properties underlying collaborative-reflection hijacking: reflective persistence and cross-agent propagation. Then building on these findings, we propose VirusCascade, the first black-box targeted promotion attack that jointly shapes semantic and structural attack surfaces: the former ensures the target item is naturally rationalised as satisfying broad user preferences, the latter positions it for system-wide propagation. Extensive experiments on four real-world datasets across diverse LLM-ARS architectures demonstrate that VirusCascade consistently achieves state-of-the-art targeted exposure under evaluated stealth constraints, reaching a mean E@20 of 0.384 and surpassing the strongest baseline by an absolute margin of +0.185.

Coding agents powered by large language models (LLMs) are evolving from making localized code changes to developing complete software repositories. However, evaluating repository-scale generation remains challenging: tasks must demand system-level reasoning while ensuring that all evaluated behaviors are precisely specified and independent of any particular implementation. We introduce E2E-SWE, a benchmark for evaluating whether coding agents can build complete, functional software repositories end to end. E2E-SWE contains 186 whole-repository generation tasks spanning 11 programming languages. Given only a natural-language specification and an empty workspace, an agent must implement a complete, installable project that satisfies a comprehensive suite of hidden tests. Each task is constructed by a software engineer in collaboration with an LLM; together, they develop the test suite and a corresponding implementation-independent specification. To ensure that tasks are well specified and practically solvable, we further subject them to an iterative verification process in which autonomous agents audit and repair task defects using static inspection and failures observed from real model rollouts. Evaluating 13 frontier models, we find substantial variation in end-to-end repository generation ability, with pass@1 ranging from 11.7% to 67.7%, providing strong model differentiation while leaving considerable headroom for future progress. Analysis of agent trajectories further reveals long, front-loaded reasoning patterns, highlighting the planning and system-level reasoning required to construct working codebases from scratch.

Large language models (LLMs) are increasingly deployed as agents for multi-step decision-making, yet transfer poorly to unseen environments. World-model methods address this by training agents to predict future observations, at the cost of additional training and errors that compound when predictions are used for planning. However, for LLM agents operating in digital environments, much of this world knowledge is already internalized during pretraining, which shifts the problem from acquiring it to eliciting it. We argue that typical post-training provides little pressure for such elicitation, since supervision under a single goal at each visited state inadvertently drives policies to rely on superficial contextual habits. We introduce EVOKE, a post-training method that supplies this pressure through goal diversity at fixed states. Motivated by theory showing that an agent competent across diverse goals must encode a world model recoverable from its action preferences, EVOKE holds the environment state and interaction history fixed and ranks the same candidate actions under alternative goals, forcing action preferences to change, so that a policy relying on contextual habits or single-goal correlations cannot order them correctly. This implicitly elicits the policy's pretrained world knowledge to inform decisions. We evaluate EVOKE across diverse tasks in three backbones, demonstrating improved task performance, unseen environment generalization, and data efficiency. We further conduct controlled analyses to better understand what drives these gains. These findings offer a new perspective on eliciting internalized world knowledge for transferable action through direct decision supervision.

Modern agentic systems combine an AI model with a harness that controls execution and environmental interactions. Harness design strongly affects long-horizon performance, yet its combinatorial search space demands substantial human effort that must be repeated as models change. Existing automated methods explore this space narrowly, optimizing only components such as prompts or skills or becoming trapped by fixed, exploitative search strategies. We introduce MILO (Meta-evolutionary Island Orchestration), a framework that co-evolves agent harnesses and the strategy used to discover them. MILO combines: (i) hierarchical lineage memory over island-based trees, using rejected mutations as negative evidence; (ii) per-island mutator agents that rewrite complete harnesses using global search history and parent-specific feedback; and (iii) an orchestrator that adapts search through lineage grafting and speciation, mutator reassignment and curriculum revision. Across Terminal-Bench 2.1, PaperBench, and DeepSWE, MILO-discovered harnesses outperform eight state-of-the-art harnesses and six search methods using frontier (Opus 4.8) and open-weight (gpt-oss-120b) models. With Opus 4.8, MILO improves resolution over its initial harness by $+12.0%$, $+28.3%$, and $+10.3%$, respectively, compared with best prior-search gains of $+4.5%$, $+18.3%$, and $0%$. On Terminal-Bench 2.1, it achieves $86.1 \pm 2.0%$, exceeding the official leaderboard's top entry ($83.8 \pm 2.3%$) while using 26% fewer tokens than its initial harness. On EinsteinArena open problems, MILO improves best-known upper bounds for Erdős minimum-overlap ($0.3808586 \to 0.3808568$) and the first and third autocorrelation inequalities ($1.50274365 \to 1.50274360$; $1.45081 \to 1.44889$).

Multi-agent systems of LLMs add discussion to majority voting and are therefore expected to be more capable. However, empirical reports conflict on whether discussion improves accuracy or leads to an incorrect consensus. Here, we introduce a parsimonious model that explains when discussion improves accuracy and when it ends in an incorrect consensus, built from four behaviors repeatedly observed in LLM agents: (1) withholding dissent, (2) internalizing a stated answer, (3) reconsidering after seeing dissent, and (4) correcting toward the correct answer. The model shows that discussion can overturn an incorrect initial majority only when the withholding rate $c$ is below a critical rate $c^* = γ/(γ+ a)$, set by the net correction rate $γ$ and the internalization rate $a$. We estimate these rates from conversation logs with a Bayesian method and place LLM teams relative to $c^*$. As the model predicts, the gain from discussion shrinks as withholding rises, across LLMs and on a hidden profile benchmark, HiddenBench, and MedEInst. Instructing agents not to withhold dissent increases this gain. Turning reasoning off also increases the gain, because reasoning raises the internalization rate $a$ and keeps agents from reconsidering a minority answer. These findings reconcile the conflicting reports and identify when discussion outperforms majority voting.

Existing architectures for LLM-based multi-agent systems (MAS) cannot reliably and efficiently solve multi-step tasks at scale: they struggle to support large numbers of agents and concurrent tasks, tolerate failures, govern agent interactions, and accommodate the diverse planning and execution patterns different tasks require. We present PANDA, a decentralized architecture that connects a large collective of heterogeneous, independently administered agents, letting them discover each other's capabilities and self-organize into small specialized teams per task. PANDA scales by decoupling collective communication from team communication, allowing agents to participate in multiple teams simultaneously, load-balancing tasks across the collective, and scheduling concurrent work within each agent. PANDA further separates the underlying architecture from the orchestration strategy, supporting three planning and execution patterns (star, chain, and mesh) that can be selected according to the structure and requirements of each task. PANDA detects infrastructure and orchestration failures and recovers affected tasks by dynamically replanning around failed components. Finally, to provide governance without a centralized service that would limit scalability, PANDA uses a web-of-trust model to constrain agent interactions to established trust relationships. We evaluate PANDA on the HotPotQA benchmark, demonstrating that it scales to thousands of agents, assembles teams in milliseconds, matches state-of-the-art accuracy at up to 8x the efficiency, and sustains 100% task completion under faults where existing systems fail.

Pragma insertion drives the quality of high-level synthesis (HLS) designs. Choosing the right directives demands expert knowledge and reasoning about loop nesting, data dependences, and memory layout. While existing large language models (LLMs) show promise in code generation, they lack explicit program-structure awareness, limiting their ability to suggest effective pragmas. We present PRISM, a novel structure-augmented LLM that closes this gap by adding compiler-grade structural reasoning to a pretrained, frozen code LLM. It combines three hierarchical program representations, Abstract Syntax Tree (AST), Control-Flow Graph (CFG), and Data-Flow Graph (DFG), injecting them into a specific transformer layer while keeping original code tokens in a separate stream. The cross-attention gate at the injection point allows falling back to the pretrained representation when its structural signal is unhelpful. On zero-shot evaluation in HLS-Eval, PRISM synthesizes 3.5\times as many kernels as Llama3-8B (26.9% vs. 7.7%), and on the kernels where it does succeed, it produces designs that are 2.31\times faster (geomean) than GPT-5-mini's. In the agentic flow, the PRISM codegen outperforms other baselines when optimizing complex code and drives the average normalized improvement across the HLS-Eval suite to 26.4%.

Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.

Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group standard deviation. The corresponding variance equals the sum of all pairwise reward covariances. For a fixed centered reward, larger aggregate covariance produces smaller advantages, and vice versa, allowing update magnitudes to adapt to reward dependence. However, correlated rewards with large scales can dominate this normalization and suppress signals from smaller-scale rewards. We propose Correlation-Normalized GRPO (CorrGRPO), which normalizes pairwise covariances into Pearson correlation coefficients. CorrGRPO keeps the centered total reward unchanged while balancing the influence of differently scaled rewards on the correlation-based normalization. This allows advantage magnitudes to adapt to reward correlations without the normalization being dominated by large-scale reward components. We compare CorrGRPO with GRPO and other variants on code generation, tool calling, and agent security, using models ranging from 0.5B to 8B parameters. These tasks all involve multiple rewards that can improve together or present tradeoffs. Results show improvements across three domains, including code generation, tool calling, and agent security. Our code is available at https://github.com/HKUST-KnowComp/CorrGRPO.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.

Visual agents solve problems by interleaving reasoning with image operations, and on-policy distillation (OPD) provides guidance from a strong teacher on student-generated interaction trajectories. However, image operations change the evidence available for subsequent reasoning, so local errors in evidence acquisition (Acquire), reading (Read), or answer grounding (Ground) can propagate through the trajectory and lead to incorrect answers. Existing multimodal OPD methods primarily construct or contrast auxiliary views of the original image to strengthen supervision, without explicitly modeling the connections between student actions, resulting observations, and subsequent reasoning. This limits their ability to provide corrections tailored to different failure stages. We introduce Reflection on Visual Evidence (ReVuE), an on-policy distillation method for visual agents. ReVuE compares multiple student-generated trajectories for the same query, summarizes the observed visual evidence, and diagnoses the first failure across the Acquire, Read, and Ground stages. The resulting reflections provide training-time context for the teacher. We group and reweight token-level distillation losses according to how strongly these reflections affect the teacher's predictions. This design translates trajectory-level evidence diagnosis into targeted token-level supervision, guiding students to improve their visual evidence acquisition and reasoning. Across 11 benchmarks spanning the Qwen2.5-VL and InternVL3.5 model families, ReVuE outperforms all evaluated OPD baselines in weighted-average scores for perception, mathematical reasoning, and general tasks. ReVuE also reduces redundancy in reasoning and tool calls while improving tool-call accuracy and task accuracy. Code is available at https://github.com/sylvain-wei/ReVuE

Frontier coding agents can now write and execute code that authors 3D environments, but whether they reliably understand 3D structure and precisely control scene state remains unclear. The generated 3D scene is a persistent, executable artifact: a convincing render can hide incorrect spatial relations, intersecting objects, or unintended modifications. We introduce Code4Scene, a benchmark of 190 Unreal Engine cases built from human-assembled scenes that evaluates coding agents on two complementary settings under a shared execution interface. Construction tests scene-level spatial reasoning from open-ended language specifications, where many realizations are valid; editing tests precise control of scene state, where the agent must recover the target scene from reference images while preserving everything else. Rather than scoring code or rendered views, Code4Scene evaluates the generated engine-native scene for task fulfillment, artifact integrity, and static physical validity, with edits additionally compared against withheld ground truth. Across 14 coding-agent configurations on the 95-case public set, construction and editing performance are strongly correlated but not interchangeable (Spearman $ρ= 0.78$): Claude Fable 5.1 leads construction, Gemini 3.8 Flash leads editing, and GPT-6 Astra narrowly leads overall. Spatial Composition is the weakest construction category for every agent, while editing remains imprecise: the best Repair F1 is only 0.527, and 35.8% of edits that fully recover the target still introduce unintended changes elsewhere in the scene. These results expose a gap between plausible 3D generation and reliable spatial reasoning and state control.

Average performance on routine driving benchmarks does not establish planner reliability under rare, safety-critical hazards. We proposed ExceptionDrive, a counterfactual planning benchmark that uses VLM-assisted screening, localized multi-view editing, and quality auditing to insert hazards into real nuScenes scenes while preserving their context. Its 21 tasks span six safety families and define hazard or conflict regions, local safety constraints, and acceptable responses. Because hazard insertion can invalidate the recorded human trajectory, our reference-free protocol evaluates edited predictions using Unsafe Rate (UR), Hazard Clearance Compliance (HCC), Hazard Proximity Response (HPR), and Counterfactual Trajectory Shift (CTS), which measure core-region intrusion, clearance compliance, clearance relative to a prescribed margin, and counterfactual trajectory change. Seven representative planners frequently intrude into hazard regions or provide insufficient clearance. We also develop a Reminder Agent that, without sample-specific task labels, converts visual evidence and the shared taxonomy into structured records of hazard presence, type, and a recommended high-level strategy. The agent neither predicts trajectories nor controls the vehicle; its records guide a VLM-based decision agent. In zero-shot experiments, the reminders improve strategy accuracy and reduce under-warning.

World-action models (WAMs) jointly predict how a scene will evolve and how an agent should act, however joint generation alone does not necessarily impose a shared geometric constraint on these predictions. We present PhysWAM, a unified world-action model for autonomous driving that co-denoises multiview video, metric depth, and ego motion within a single flow-matching transformer. To ground world and action generation in measured scene geometry, we introduce Coupled Point Projection (CPP) that unprojects the generated depth into 3D points, transforms them using the generated $\mathrm{SE}(3)$ ego motion, and minimizes their distance to LiDAR points transformed using the recorded ego motion. This geometric constraint promotes physical consistency with the measured scene by jointly supervising generated depth and motion alongside their standard flow-matching objectives. At inference, trajectory selection relies only on a simple label-free consensus rule, with no learned scorer or simulator feedback. We evaluate PhysWAM across NAVSIM v1 and v2 planning, zero-shot closed-loop transfer, and future video and metric-depth prediction. Despite PhysWAM's simple selection procedure, it achieves strong planning performance and transfers zero-shot to unseen driving environments. It also generates accurate metric depth and temporally coherent video, with CPP improving both planning and depth prediction. Together, these results demonstrate that the geometric relationship between scene depth and ego motion provides a direct way to couple world and action generation within a simple unified model.

Chain-of-thought traces are widely read as records of how models reach their answers, informing debugging, agent auditing, and claims about reasoning. Testing this interpretation is difficult because natural-language thinking traces are rarely mechanically verifiable. We revisit it in iGSM, a synthetic grade-school mathematics benchmark designed to study thinking traces and used to support claims of learned reasoning and planning. Crucially, iGSM exposes the exact quantities and dependencies that a correct solution should use, allowing generated traces to be checked programmatically step by step and enabling us to test whether correct answers are reliably accompanied by valid traces. We first evaluate models trained exclusively on valid, minimal traces. Answer correctness and trace validity nearly coincide in distribution but decouple out of distribution: on the hardest instances, 31.6% of correct answers have invalid traces, over half of which pass all syntactic and arithmetic checks but fail semantic dependency checks. We then intervene on trace supervision. Non-minimal training traces induce non-minimal outputs, while re-asking the same problem with a different query reveals computations inherited from the original query, weakening minimality as evidence of selective planning. Shuffling tokens in 10% of training trace sentences preserves near-clean accuracy even out of distribution despite no trace passing verification. Swapped training traces likewise retain high in-distribution accuracy. We discuss the implications of these findings for chain-of-thought monitoring and interpretation in the context of AI safety.

Software engineering agents increasingly use reusable skills distilled from prior experience to resolve repository-level issues, yet such skills often fail to transfer across repositories. A central challenge is that a behavior appearing in a successful trajectory is not necessarily responsible for the successful outcome: it may be genuinely useful, merely incidental, or simply a recurring habit of the model. We introduce XRepoSkill, a trajectory-based approach for learning transferable skills. We represent a skill as a collection of rules, each specifying what action to take and when to take it during issue resolution. XRepoSkill first contrasts successful and failed trajectories of the same agent on the same issue and derives candidate rules from where their execution paths diverge. Each rule is paired with an executable predicate that enables its prescribed behavior to be evaluated systematically on other trajectories. A rule is verified based on its association with successful issue resolution and retained only when its prescribed behavior recurs across multiple repositories; repository-specific variants of the same behavior are then consolidated into transferable rules. For a new issue, XRepoSkill selects relevant rules to guide the agent. We learn skills from publicly released trajectories on the official SWE-bench Verified leaderboard and evaluate them on SWE-bench Pro and DeepSWE using three backbone LLMs from different vendors; none of the evaluation repositories appears in the skill-learning trajectory pool. Against three recent skill learning methods, XRepoSkill achieves the highest issue resolution rate in all six benchmark--LLM combinations. In particular, on the challenging long-horizon DeepSWE benchmark, XRepoSkill improves issue resolution by 10.3 percentage points over the same agent without learned skills and by 5.0 points over the strongest skill-learning baseline.

As the capabilities of large language models (LLMs) continue to advance, increasing attention is turning to how to translate their abilities into useful behavior. Personal agents bring this question into everyday settings, where models are expected to serve individual users and continually adapt to their preferences. With the underlying model held fixed, such adaptation relies on harness engineering: designing and evolving the surrounding layer that manages context, memory, tools, and execution. Despite rapid progress, the factors governing effective harness evolution remain insufficiently understood. To narrow this gap, we investigate three central questions concerning harness architecture, harness scale, and self-evolution algorithms through complementary empirical and theoretical analyses. Empirically, we introduce a preference-oriented benchmark and systematically characterize the capabilities and limitations of personal agents associated with these three dimensions. Theoretically, we formulate harness evolution as a learning problem and explain these phenomena through approximation, generalization, and optimization errors. Analyses of reachable policies, capacity under finite interaction evidence, and biased update dynamics provide theoretical accounts of the observed phenomena. Together, these results offer a unified perspective on the limits of personalization through harness evolution and inform future harness design.

Skills equip LLM agents with professional knowledge and guidance to complete long-horizon and complex tasks. Although skills have been widely adopted in recent agent paradigms and harnesses, how to synthesize reliable training data and how to train agents for skill use remain underexplored. In this work, we propose SkillGym, an automatic pipeline to build verifiable environments, collect trajectories, and train skill-use agents. SkillGym first crawls a large volume of skills from the internet, then keeps those whose workflows can run reproducibly offline. A builder-reviewer pipeline is used to construct difficulty-controlled tasks, spanning four task types, each with a reference solution and an executable verifier. With this pipeline, we build 6.8k environments and collect 19k verified successful trajectories for supervised finetuning. Finetuning on these trajectories improves LLMs of different families and sizes, from 2B to 122B parameters across four skill-use benchmarks; Our Qwen3.5-9B SFT model outperforms the 397B untrained model on two of them. Further analysis shows that training teaches agents to invoke skills, raising the rate of reading the relevant skill from 28% to 96%, and that the gains hold across reasoning structures, extending to task types that form a minority of the training data and to skills held out from training

LLM agents performing long-horizon tasks accumulate tool results that later steps may need. Passing the full history to every invocation is costly even when it fits within the context window, while reducing it risks omitting needed information. Existing context management methods can overlook how earlier tool results are used in subsequent execution, leaving needed information out of context. We introduce ContextRender, which manages context through a persistent graph of execution dependencies. We develop Tool-Flow Analysis to track how later operations reuse information from earlier tool results, providing a signal called observed reuse. A renderer combines this signal with recency and semantic relevance to select results within a fixed history budget, retaining omitted results for later use. Across AppWorld and 8-objective QA with three execution models, ContextRender outperforms the evaluated context management baselines using a 6K history budget, well below the models' maximum context windows. Within this budget, it achieves task performance close to or above that of passing the full history while reducing mean inference cost by 10.2%-32.2% relative to Full history. Ablations show that observed reuse improves task performance and retention of results reused later.

Vision-language-action and world-action models have demonstrated impressive capabilities in robotics, yet generalization to unseen tasks remains challenging. More recently, general-purpose multimodal agents have shown great potential for zero-shot robotic task solving. However, they often incur high execution costs by reasoning and exploring the physical world from scratch. To reduce these costs, we introduce RoboSkill, a framework that connects skill acquisition and reuse through an Explore, Execute, Evolve loop. Within this loop, the agent explores to gather task-relevant information, executes tasks while adapting to feedback, and evolves its skill library based on execution records. It then reuses these skills to guide exploration and execution in the next cycle, closing the loop. To improve loop efficiency, we complement vision with tactile feedback to reduce uncertainty during physical interaction. We further augment textual guidance with reusable code to reduce reasoning overhead during skill reuse. On LIBERO-10, RoboSkill improves first-episode success rates by 12.5--25.0 percentage points and reduces average runtime by 7.6--72.4% across four agents. On real robots, it improves success rates by 8.3 percentage points and reduces average runtime for successful trials by at least 14.4%.

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Agentic reinforcement learning (ARL) with verifiable rewards improves the ability of large language models (LLMs) to tackle knowledge-intensive tasks by learning to interleave search and reasoning. However, most existing ARL methods optimize only LLM-generated tokens and treat retrieved evidence as environment observations. This creates an information-credit gap: failures caused by missing or misleading evidence are attributed to the LLM policy rather than to the retriever, which motivates training the LLM and the retriever jointly. In this paper, we show that retrieval and LLM policy learning are order-sensitive: adapting the retriever before optimizing the policy yields a larger reward gain than the reverse order. To preserve this hierarchy while allowing both components to co-adapt, we formulate retrieval-augmented agentic RL as a bilevel optimization problem. To solve it efficiently, we introduce BRIDGE, a memory-efficient first-order bilevel method motivated by a loss-landscape analysis of the RL and retrieval objectives. Across seven open-domain QA benchmarks, BRIDGE achieves the highest average accuracy with both 3B and 7B backbones, improving the multi-hop average over the strongest baseline by 9.6 and 3.4 EM points, respectively. It also achieves the best averaged answer accuracy and reasoning quality across medical QA benchmarks.

Large language model (LLM) agents reuse external memory to guide new tasks, but effective retrieval requires learning which memory sets improve execution. Such learning relies on costly outcome feedback: ordinary retrieval observes only executed sets, while evaluating alternatives requires additional rollouts. We introduce \textsc{UpliftMem}, which learns memory retrieval from set-level execution uplift relative to the same executor without memory. A theoretical analysis of how retrieval preferences restrict feedback coverage motivates targeted probing of alternative memory sets. Probe selection follows an expected value of sample information (EVSI) criterion, derived in closed form under a correlated Gaussian model, to allocate limited training rollouts according to their expected improvement in local retrieval decisions. The shared scorer is trained with a frozen executor and selects memory sets without test-time probes. Across ALFWorld, WebShop, and BigCodeBench, \textsc{UpliftMem} achieves the best success rates among evaluated baselines on the main evaluation sets. Controlled fixed-store and matched probe budget evaluations further demonstrate improved memory-use decisions and more effective use of execution feedback.

LLM agents accumulate interaction histories that grow linearly with task length, causing quadratic inference cost scaling and performance degradation from attention dilution. Existing context-compression methods learn what to discard offline: by contrastively optimizing guidelines, distilling compressors, or training compression policies. This incurs a substantial cost. Further, the compression policy is learned a priori and is not dynamically conditioned on the evolving test-time trajectories. In this paper we ask a complementary question: Which past interactions causally shape the agent's future decisions? We recast context compression as a causal decision preservation problem over discrete interaction units and introduce FOCUS, a training-free context compression framework that operates entirely at test time. Our method requires no offline data collection or fine-tuning, and is architecture-agnostic, attaching to any closed-API frontier model as a modular compression layer. We evaluate FOCUS on diverse agentic benchmarks including API and tool-calling, QA, web domain and multi-turn dialogue. Our method establishes new state of the art performance, cutting peak context by up to 48% and dependency by 73% while improving task success by up to 8.9 percentage points over uncompressed execution.

Autonomous agents have made rapid progress in general-purpose computer use, yet reliable automation of professional industrial engineering remains out of reach. Engineering workflows require reasoning over geometric and physical constraints and keeping dependencies consistent across software and design stages. We present EngiWorld, a benchmark designed around the capabilities that real engineering projects require, from operating a single application to solving engineering problems end to end. It contains 1,301 expert-curated tasks spanning six engineering domains (CAD, CAE, CAM, BIM, EDA, and DCC) and 26 professional software applications, with both GUI and CLI interfaces. The tasks are organized into six types, corresponding to capabilities that real engineering projects require. We further introduce an artifact-centric evaluation methodology built on a unified suite of domain verifiers. The verifiers programmatically check whether final and intermediate artifacts are geometrically valid, physically feasible, and compliant with design rules. Beyond binary success, they also score how well optimized designs meet their specifications. Evaluation of seven frontier models reveals a substantial capability gap: the strongest model scores only 44.7/100, and just 3.6% of runs on tasks that require cross-software coordination succeed. EngiWorld provides a rigorous foundation for measuring progress toward agents that operate professional engineering software end to end.

Active multimodal agents use visual tools to acquire task-relevant evidence while reasoning. Although reinforcement learning samples multiple interaction trajectories per input, outcome-based objectives primarily use the group to estimate scalar advantages, leaving complementary visual discoveries underused. We introduce VISTA, which internalizes collective visual experience through on-policy distillation by turning observations from same-input rollouts into shared supervision. Collective visual experience distillation (CVED) organizes these observations with their interaction context and aligns them with individual decisions, while heterogeneity-aware policy improvement (HAPI) reinforces successful trajectories and provides experience-guided distillation for unsuccessful attempts. An experience-conditioned teacher evaluates the student's sampled response prefixes, allowing discoveries from one trajectory to guide learning in another without replacing the student's original history or generating new target trajectories. The trained agent retains its visual tools and acts using its own interaction history. VISTA achieves the strongest average performance among the evaluated active multimodal agents of comparable size and consistently outperforms same-backbone training baselines across fine-grained perception and general reasoning tasks, demonstrating the value of collective experience for active multimodal learning.

As large language models (LLMs) advance, AI agents are increasingly deployed in open-world environments to tackle complex sequential tasks (e.g., document processing, cross-application collaboration), relying heavily on actions ranging from GUI operations to semantic APIs. However, three core challenges persist: the "scale dilemma" of massive tool ecosystems exceeding LLM context windows, the "non-stationarity" of tool quality due to updates or outages, and the "heterogeneity" of feedback formats (pixels, text, structured data) creating information silos. To address these, we propose AnyAct, a universal action layer that unifies available capabilities into a self-evolving action space, enabling agents to operate efficiently and reliably in large-scale, dynamic tool ecosystems. AnyAct's core design focuses on two objectives: constructing this action space via hierarchical progressive retrieval (filtering task-relevant actions) and test-time reliability evolution (pruning unreliable actions), and enabling reliability-aware action orchestration through a heterogeneous observation grounding module that unifies multi-modal feedback. Additionally, it defines a hybrid action space (primitive + semantic actions) and optimizes for a balance between task success rate and execution cost. Evaluations on LiveMCPBench and OSMCP (a new benchmark we developed for multi-granularity action collaboration) demonstrate state-of-the-art performance. AnyAct delivers substantial performance gains over baseline methods across various LLM base models on LiveMCPBench and improvements are particularly notable for models with constrained native capabilities. On OSMCP, it achieves 77.27% overall success with only 50 steps, which is half the steps required by most competitors.

Long-form video understanding requires multimodal agents to iteratively gather evidence over many reasoning steps. However, most existing agentic methods suffer from semantic thrashing: as append-only working memory grows, attention to key evidence collapses, and the agent loses access to what it has already found. First, we provide a structural argument showing that append-only memory can incorporate newly observed target evidence, but cannot remove accumulated noise or prevent ordered context growth without a rewrite operator. Second, motivated by this analysis, we propose VideoLoop, a multimodal agent with two coupled loops. The outer loop reasons over the video and the inner loop, after each step, retrieves artifacts from an unbounded filesystem of past observations and intermediate analysis, and rewrites a bounded working memory. Extensive experiments demonstrate the effectiveness of VideoLoop, which improves four popular LVLM backbones in a plug-and-play manner, with an average gain of 4.2% points over baseline on VideoMME (long). Further analysis of working memory suggests that VideoLoop mitigates semantic thrashing: on the hardest quarter of VideoMME (long) questions, a blind judge that reads only the agent's context answers 81.1% correctly, versus 60.9% for the append-only agent. With Gemini 3.1 Pro, VideoLoop reaches 88.3% on VideoMME (long), 88.8% on VideoMMMU, and 80.9% on LongVideoBench (long).