Agent 进展

As LLM agents begin to take real, irreversible actions (shell commands, file edits, deploys), the standard safety pattern is a human-in-the-loop approval gate: risky actions pause and wait for a person. We argue the gate is the easy part; the hard part is the judgment - which actions to stop - which the field evaluates against two false assumptions: that there is a ground-truth notion of "risky," and that the human reviewer is a perfect, infinitely-available oracle. On a hand-labeled set of 125 adversarially-weighted agent actions we show that (i) reviewers only moderately agree on what is risky (Fleiss' kappa = 0.52), so there is no single correct label; (ii) framing the guard as selective classification under asymmetric cost makes its operating limits measurable, and on hard inputs the guard cannot safely auto-decide; and (iii) when the reviewer is modeled as endogenous (fatiguing as escalation load grows), realized safety becomes an inverted-U in the escalation rate: more human oversight can make a system less safe, and the safety-optimal guard escalates below full escalation - a setting a load-aware policy also uses to resist a flooding attack that slips a malicious action past a fatigued reviewer. Agent oversight, framed this way, is not only a classification problem but a resource-allocation one: human attention is finite, and the guard's escalation policy spends it. We claim none of these mechanisms as novel - fatigue-aware learning-to-defer (FALCON), cost-sensitive deferral under workload constraints (DeCCaF), trajectory-level guarding, and reviewer-fatigue/flooding attacks are all prior art we cite. Our contribution is an open-source agent-oversight system that operationalizes and measures them in the LLM-agent action-gating setting, turning "is my guard good?" from a guess into a curve. The inverted-U and the flooding attack are modeling results that motivate a human study.

Tool-using LLM agents interact with the world through actions that persist state in artifacts (e.g., workspace files or logs). Consequently, jailbreak defenses must reason about cross-step composition rather than isolated text. Yet most existing attacks and defenses, including ``multi-turn'' jailbreaks such as Crescendo and Tree of Attacks,still assume a single contiguous conversation visible to the defender. This assumption breaks down in real agent pipelines, where enforcement is fragmented across tools, modules, and time, and where artifact provenance is often not tracked. We operationalize a deployment failure mode for tool-using LLM agents, the provenance gap, and study reproducible triggers for it: Context-Fractured Decomposition (CFD), a family of cross-context multi-step jailbreaks that preserve benign-looking intermediate artifacts from an early interaction and elicit harmful behavior much later, potentially in a different agent instance or workflow stage, via individually innocuous tool actions whose risk emerges only under delayed artifact-mediated composition. We instrument the failure mode with trace-level diagnostics and outline a verifiable mitigation direction (provenance lineage tagging). Across agent-system jailbreak benchmarks, CFD improves success rates by up to 28.3 percentage points over state-of-the-art baselines, even against strong single-turn judges. Disclaimer: This paper contains examples of harmful or offensive language.

We present AliyunConsoleAgent, a web agent framework for automated documentation verification in real-world cloud consoles. Major cloud platforms encompass hundreds of products with rapid feature iteration, causing console UIs to frequently diverge from their corresponding documentation. Verifying that documented procedures accurately reflect the current console and can be executed end-to-end demands an estimated 4 million recurring inspections annually, yet manual coverage remains below 1%. While agent systems built on frontier proprietary models achieve high success rates, their prohibitive cost and data privacy constraints preclude large-scale deployment. We propose a two-stage training paradigm: supervised fine-tuning (SFT) on distilled frontier-model trajectories, followed by reinforcement learning using Group Relative Policy Optimization (GRPO) and a dual-channel outcome reward model in real cloud environments. To support large-scale RL training, we construct a high-determinism rollout system featuring Terraform-based resource pre-provisioning and LLM-driven on-demand provisioning, which effectively isolates environment noise from the training signal. We further introduce a rule-based reward evaluation protocol grounded in backend audit logs, providing objective, reward-hacking-resistant outcome judgment. Our model evolves from mechanical instruction following to autonomous decision-making with cloud console and product-specific understanding. Experiments on a challenging 278-task benchmark where the best frontier model achieves only 65.34% demonstrate that AliyunConsoleAgent-32B achieves a 63.52% mean success rate — a 20.24 percentage-point improvement over the base model, narrowing the gap to the best frontier proprietary model to 1.82 pp (bootstrap 95% CI [-1.27, 7.39]) — at 92% lower inference cost.

Tool-using large language model (LLM) agents face two distinct security failures: unauthorized external actions and exposure of sensitive plaintext inside the runtime before any final output check can intervene. Existing defenses usually protect one boundary, either the planner/runtime or the action sink, and therefore do not by themselves secure both surfaces. We present SecureClaw, a dual-boundary architecture that places authorization at the effect sink and plaintext confinement at the read boundary. Sensitive reads pass through a trusted gateway that replaces raw values with opaque handles and, in the evaluated deployment, bounded summaries as an explicit declassification interface. Writes that change external state follow a PREVIEW$\rightarrow$COMMIT protocol in which only a trusted executor may commit the exact canonical request authorized by policy. The runtime can still plan over summaries and symbolic references, but cannot directly dereference secrets or perform side effects. Across AgentDojo, AgentLeak, and Agent Security Bench (ASB), SecureClaw is the only defense we evaluate in a common harness that simultaneously retains usable task utility and achieves 0% attack success rate (ASR) on ASB, 0.64% ASR on AgentDojo, and 3.23% overall leak on AgentLeak's attacked parity lane, which measures final-output and internal-relay leakage.

The dominant paradigm in video retrieval relies on embedding-based full-corpus scanning, which suffers from inherent computational inefficiency and the semantic asymmetry between information-dense videos and sparse textual queries. To bridge this gap, we introduce MAVIS, a novel multi-agent framework that rethinks retrieval as cooperative reasoning rather than brute-force search. MAVIS first bridges the granularity mismatch by parsing raw videos into a Structured Semantic Library, enabling explicit attribute-level indexing. During retrieval, a planner decomposes complex user intents into atomic sub-tasks, dispatching specialized agents to independently nominate candidates. Crucially, MAVIS employs a Logic-aware Debate mechanism with a strict veto protocol, where agents collaboratively prune logical mismatches to identify a compact set of ``controversial'' candidates for fine-grained verification. This agentic workflow effectively bypasses the inefficiency of full-library traversal. Extensive experiments on MSR-VTT, MSVD, and ActivityNet demonstrate that MAVIS achieves competitive performance without task-specific fine-tuning, offering a scalable and interpretable alternative to traditional dual-encoder approaches.

Vision-language model (VLM) agents are increasingly deployed in interactive game environments. Yet game benchmarks for VLM agents typically report a single first-attempt score per (agent, game) pair, focus on single-agent Solo play, and lack unified protocols for evaluating heterogeneous agent classes (commercial VLMs, open-weight VLMs, and specialized game policies) on the same footing. We address these gaps with OmniGameArena, a real-time benchmark of twelve newly built Unreal Engine 5 games spanning Solo (7), PvP (3), and Coop (2) with unified action interfaces, and the Improvement Dynamics Curve (IDC), an agentic-reflection harness in which a tool-using reflector LLM autonomously refines a bounded skill prompt across multiple rounds. Beyond cold-start leaderboard scores, IDC exposes two additional observables for each (agent, game) pair: how the score evolves across reflection rounds, and how the learned skill behaves on held-out task variants. We report these observables for twelve VLM agents on the cold-start leaderboard and four top agents under IDC.

Text-driven indoor scene generation and editing require an intermediate representation that language models can both produce and revise. Existing LLM-based systems often rely on scene graphs or global constraint lists, which are compact but underspecify local geometry and make instruction-based edits difficult to localize. We frame this problem as structured program generation and local program repair, and propose Hierarchical Descriptive Scene Language (HDSL), an XML/CSS-style domain-specific language for structured 3D indoor scenes. HDSL represents rooms, regions, objects, and support surfaces as a tree with local coordinates, making complex scenes easier to plan recursively and easier to retrieve for editing. Our pipeline uses LLM agents to generate HDSL subtrees with bounded verification, grounds non-virtual nodes through multimodal asset retrieval, and applies force-directed layout optimization to repair boundary and collision errors. For editing, Hierarchical Retrieval-Augmented Generation retrieves the relevant subtree, asks the LLM to rewrite only that local context, and merges the result back through a deterministic three-way merge. In our reproduced benchmark, HDSL improves average object coverage, text-scene alignment, and generation time over full text-to-scene baselines while remaining competitive with recent layout-only reproductions on geometry metrics; for editing, HRAG reduces token use by $5.22\times$ and runtime by $6.19\times$, produces valid DSL for all eight paired edits, and better preserves unrelated scene objects.

Medical agent systems are increasingly expected to support interactive clinical decision making rather than only static question answering. In such settings, effective agents must reuse prior experience across evolving cases, yet existing memory mechanisms often retain raw historical traces that are redundant, noisy, and difficult to govern. More importantly, they rarely distinguish which memories are truly useful for future reasoning. This limits their ability to accumulate compact and reliable experience for long-horizon clinical reasoning. To close this gap, we propose SkeMex, a post-deployment self-evolution framework that improves medical agents through a skill-based memory without updating model weights. SkeMex distills informative interaction trajectories into structured skills that encode reusable procedural knowledge, and organizes them into a multi-branch repository spanning general, task-specific, and action-level experience. To determine which memories should be reused and retained, SkeMex estimates context-dependent utility from environment feedback and uses it to guide value-aware retrieval and repository governance. A closed-loop ``Read--Write--Assess--Govern" lifecycle further supports continual evolution by writing new skills, updating utilities, promoting useful memories, and removing harmful entries. Experiments across diverse clinical tasks show that SkeMex consistently outperforms representative memory-based agents in both offline and online settings. It also generalizes across model backbones and supports transferable skill memory. All data and code will be released publicly.

Long-term memory for an LLM agent is more than retrieving the right passage at the right time. Current memory systems collapse belief revision, causal coupling, and cross-domain abstraction into a single retrieval surface tuned for surface recall, and consequently struggle on implicit personalisation that requires reasoning over how a user has evolved. We propose DCPM, which reorganises agent memory along a cognitive capability hierarchy ascending from raw inputs and atomic facts, through diachronic belief trajectories and identity, to domain schemas, latent intentions and cross-domain patterns. The hierarchy is driven by two processes inheriting the architectural split of dual-process theory: a synchronous daytime writer (System1) that records belief revisions as doubly linked supersedes chains, and an asynchronous nighttime engine (System2) that induces schemas and intentions and sweeps for cross-domain collisions abstracted into higher-level core schemas. On LongMemEval, PersonaMem and PersonaMem-v2, enabling System2 contributes most where the benchmark rewards implicit cross-session inference (up to +5.20 on PersonaMem-v2) and least on span recall, matching the architectural prediction.

Vision-and-Language Navigation in continuous environments requires agents to understand the spatial structure of previously unseen environments in order to follow language instructions. Although foundation models have opened a promising path toward zero-shot navigation without task-specific policy training, many navigators still rely on local visual cues and linear history-based reasoning, overlooking the spatial nature of navigation across explored regions, traversed paths, landmarks, and their spatial relations. In this paper, we propose SpaceVLN, a navigation agent built around Spatial Cognitive Memory and Task-Guided Spatial Reasoning. Specifically, SpaceVLN introduces an efficient stagewise closed-loop framework where planning and execution are organized around verifiable space--landmark stages. During navigation, the agent progressively abstracts explored regions into Spatial Waypoints and dynamically maintains subtask-grounded landmark evidence, forming a hierarchical Spatial Cognitive Memory for progress localization and spatial-relation understanding. Built on this memory, Spatial-CoT integrates task-progress reasoning with spatial perception, analysis, and prediction, enabling Task-Guided Spatial Reasoning for embodied navigation. The unified stage interface enables SpaceVLN to address both Vision-and-Language Navigation and Object-Goal Navigation under a unified zero-shot setting, without task-specific policy training. Across R2R-CE, RxR-CE, GN-Bench, and HM3D-OVON, SpaceVLN achieves state-of-the-art zero-shot performance, and real-robot deployment further validates its applicability. These results highlight Spatial Cognitive Memory and Task-Guided Spatial Reasoning as a practical foundation for stronger embodied navigation agents.

Strabismus is a common ocular disorder that requires fine-grained subtype diagnosis for individualized treatment planning. However, existing deep learning methods mainly provide diagnostic predictions without transparent reasoning, while recent large vision-language models (LVLMs), although promising for joint image understanding and report generation, remain highly prone to hallucination in this evidence-sensitive and rule-driven medical task. To address these challenges, we propose MAGIS, an evidence-based Multi-AGent reasoning for Interpretable Strabismus diagnosis framework. MAGIS transforms black-box end-to-end generation into a structured diagnostic process consisting of candidate hypothesis generation, dual-evidence constrained context, evidence-based corrective verification, and report generation. Specifically, we introduce a Dual-Evidence Constrained Context (DECC) mechanism that jointly organizes visual evidence from the photograph of the nine cardinal positions of gaze and evidence-based clinical diagnostic rules into a constrained context for reliable diagnostic reasoning. We further develop an Evidence-Based Corrective Verification (EBCV) mechanism that verifies whether the current diagnostic hypothesis is supported by visual evidence, heatmap-based visual cues, and evidence-based clinical diagnostic rules. Hypothesis refinement is triggered when inconsistency is detected. Experiments on a fine-grained strabismus benchmark demonstrate that MAGIS not only significantly outperforms other state-of-the-art diagnostic systems, improving the weighted F1 score from 72.0% to 91.3%, but also substantially improves the clinical reliability (consistency, alignment, and completeness) of generated diagnostic reports. These results demonstrate that MAGIS provides an effective solution for building accurate, evidence-based, and clinically interpretable strabismus diagnosis systems.

Large language model (LLM) agents now solve complex tasks through long plan-and-execution traces, yet the ability to locate errors in a completed traces still lags far behind, especially in the silent failure regime. Existing approaches predict suspect steps via classifiers or LLM judges, or recover correct answers via retry, but none feed the intervention outcome back to refine the attribution itself. We propose \methodname, a method that closes this gap by diagnosing a candidate error step, testing it through controlled replay with a diagnosis-specific patch, and using the verified outcome flip as contrastive evidence to refine the final attribution. Across four localization benchmarks spanning multi-hop reasoning across domains, \methodname achieves the highest localization accuracy among same-auditor methods across all four benchmarks, with the largest gains on structured tool-use traces, while providing actionable localization even when ground-truth answers are unavailable.

Large language model agents are increasingly deployed in human-human interaction settings, such as meeting assistants and clinical documentation systems, where they must observe conversations and retain information for downstream queries. Unlike traditional human-assistant settings, these environments are inherently multimodal, involve complex discourse phenomena such as anaphora and deixis, and contain asynchronous or conflicting information from multiple participants. However, existing memory benchmarks largely focus on single-user, text-only interactions, failing to capture these challenges. To address this gap, we introduce H2HMem, a Human-to-Human Multimodal Memory Benchmark for evaluating memory capabilities in complex human-human interactions. H2HMem includes both dyadic and multi-party conversations with multimodal information streams, and evaluates agents along three dimensions: memory recall, reasoning, and application. Experiments with advanced agents reveal substantial limitations in constructing, retaining, and utilizing memories across modalities, participants, and sessions, highlighting substantial room for improvement in next-generation LLM agents.

Large language model (LLM)-based agents are increasingly used in interactive textual environments, from web navigation and code editing to tool use and long-horizon dialogue. Yet many remain largely reactive, mapping observations to actions without an explicit model of how these environments are structured and evolve. This motivates text world models (TWMs): transition models over textual states that, given a state and a candidate action, predict the resulting webpage, terminal output, API response, or user reply, thereby supporting planning, efficient learning, and principled evaluation. We systematically review text world models for LLM-based agents, organized around a formal framework and the agent lifecycle: (1) Foundations, defining text world models and characterizing them by state representation and grounding domain; (2) Construction, taxonomizing LLM-as-WM and code-as-WM paradigms and reviewing methods for building them; (3) Application, examining how world models support agents at training time through experience synthesis and at inference time through planning, verification, and adaptation; and (4) Evaluation, covering both evaluation of the world model itself and its use as an evaluation environment for agents. We aim to consolidate this rapidly developing area, clarify its design space, and highlight open challenges for future research.

Video world models have made rapid progress in generating controllable visual experiences, but most of them still simulate the world from a single observer. Extending such models to multiple agents raises a central challenge: if each agent's future state is generated independently, overlapping views may instantiate different versions of the same scene, leading to inconsistent objects, layouts, and appearances across agents. Conventional camera conditioning controls individual trajectories, but it does not explicitly couple the generation of views that should agree under shared scene geometry. We introduce Prisma-World, a camera-controllable multi-agent world model that formulates multi-agent generation as a joint geometry-aware denoising process for cross-view consistency. Prisma-World processes all agent videos within one full-attention sequence, uses a multi-agent RoPE design to distinguish agent identities while preserving synchronized temporal coordinates, and injects relative camera geometry into attention to bias overlapping viewpoints toward shared scene evidence. To further strengthen multi-view consistency and enhance global spatial perception, we augment our framework with an overlap-decaying curriculum training paradigm alongside minimap-conditioned structural guidance. To facilitate the training and evaluation of multi-agent models, we introduce PrismaDataset, a large-scale UE5 dataset with panoramic acquisition across diverse scenes, composable multi-agent view groups with flexible agent counts and complex camera trajectories, and precise camera/action annotations for consistency training and evaluation. Experiments show that a single Prisma-World model can generate high-fidelity multi-agent videos with flexible agent numbers, camera controllability, improved cross-view consistency, and spatial grounding under minimap guidance.

Optimizing large language models (LLMs) for long-horizon caregiver agents requires balancing delayed task objectives with immediate environment dynamics, such as patient distress and resistance. In dementia care, this balance is especially difficult: trajectory level rewards are too sparse for turn level credit assignment, while external LLM-based evaluators are costly and can misread fragmented or indirect patient responses. To address this issue, we propose Turn-Trajectory Group Relative Policy Optimization (\textbf{T$^{2}$-GRPO}), a framework that decouples caregiver RL into two normalized reward horizons and enforces safety through a binary hard veto. $T^2$-GRPO derives dense turn-level rewards directly from environment state transitions, measuring changes in patient distress and resistance from a frozen dementia patient simulator. These environment-grounded rewards are combined with trajectory-level evaluations through independent centered-rank normalization, which preserves heterogeneous reward signals and mitigates reward collapse. Extensive experiments on dementia caregivers show that T $^{2}$-GRPO outperforms competitive baselines, indicating a substantial improvement for emotionally sensitive caregiver scenarios that effectively handles immediate patient feedback, long-term care outcomes, and safety constraints.

Software engineering agents (SWE agents) increasingly work through tool-mediated trajectories in real repositories, yet their behavior remains difficult to characterize in concrete, observable terms. These trajectories record tool use, intermediate reasoning, evidence selection, and self-directed stopping, but they do not by themselves explain why particular moves were chosen, what evidence was trusted, or when understanding was judged sufficient. This tension makes trajectory data both limited and valuable: faithful, replayable traces can become an empirical substrate for studying agent behavior when interpreted through disciplined observation. We introduce Ada, a scoped apparatus for repository-level code understanding. Ada enters real codebases through a bounded tool interface, allowing open-ended exploration to remain recordable as finite trajectories. Across this wild-but-bounded setting, Ada chooses where to look, what to read closely, when to consolidate partial understanding, and when to close its account of the repository. We project Ada's think-action chains through observation lenses that make navigation, evidence selection, synthesis, grounding, and stopping visible without reducing behavior to raw tool counts or speculating about hidden intent. Read together, these lenses produce behavioral profiles grounded in recorded movement through software worlds. Across 408 trajectories, spanning multiple models, repositories, task families, and launch conditions, the study shows how faithful digital traces can be transformed into disciplined, comparable projections of emerging SWE-agent mindset. The results expose differences in efficiency, trajectory diversity, epistemic grounding, and the limits of intervention, while providing a methodological foundation for observing SWE agent behavior in real codebases.

AI agents increasingly take consequential actions — shell commands, cloud operations, and arbitrary tool-calls — so a trust layer must decide, per action, whether to allow, warn, block, or escalate. We argue that the right way to reason about such a layer is by threat type. Lexical (fixed-signature) threats, where danger lives in a stable token, are decidable by deterministic rules; semantic (intent-dependent) threats, where a benign and a malicious action share the same surface, are out of reach for rules by construction. We make this concrete with a negative proof: a determined, hand-authored cloud rule pack lifts held-out accuracy only 48 to 56% overall and moves the semantic categories by 0pp (data_db 29 to 29, observability 59 to 59, supply_chain 50 to 50), while a strong LLM judge carries exactly those categories. We give the judge a self-learning capability: on a corpus that is mainly semantic attacks it nearly doubles rule accuracy (48% to 83.6-85.2%) with near-zero false-blocks, and this holds across two model providers. We turn this into a self-improving dual-store system: the judge distills a growing deterministic rule floor on lexical threats (cheaper over time) and feeds a guarded RAG memory on semantic threats (a verdict-cache fails — surface-twins collapse to ~58% — so a corroboration guard lifts semantic accuracy +13pp, 70 to 84). The result is what sets AgentTrust v2 apart from its static v1 predecessor: a trust layer that self-evolves from its own stream of decisions — cheaper on the lexical class (it distils its own rules) and smarter on the semantic class (it accrues guarded precedent), while never hard-blocking a benign action. An end-to-end online replay shows the judge-call rate falling (50% to 44%) and judge-domain accuracy rising (71% to 80%), with 0 benign hard-blocks across 45,000 actions.

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce VESTA, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions, VESTA instantiaes abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.

Data agents integrate LLM-driven reasoning with relational data access, executable analytical tools, and multi-step workflow orchestration, making them increasingly central to enterprise analytics. This integration introduces new security vulnerabilities across data resources, database execution, and agent reasoning, recombining concerns from database security and general-purpose LLM-agent security into failure modes that neither line of work captures on its own. To address this gap, we present a systematic security study of data agents. Our contributions are threefold. First, we develop a layered vulnerability framework that identifies eight data agent-specific risks across interpretation, execution, and policy layers. Second, we introduce an attack taxonomy organized by adversary goal, tactic, and technique, covering three goals, seven tactics, and fourteen techniques, and pair it with an LLM-driven payload generation pipeline grounded in real database schemas. Third, we evaluate these attacks on six systems, including four open-source data agents and two production cloud analytics services. Our experiments reveal substantial security vulnerabilities across current systems and yield four key takeaways.

While data-intensive deep reinforcement learning can optimize complex control policies, scientific discovery in physical systems fundamentally requires an interpretable chain of reasoning that connects physical evidence to structured control architectures. Here, we present a self-evolving scientific-agent workflow, driven by large language models and iterative code generation, that automates controller construction while preserving strict interpretability and rigorous physical reasoning. Instead of adjusting weights, the agent deploys candidate strategies into physical simulations, actively diagnoses dynamic behaviors from multimodal evidence, and translates these observations into progressive source-code refinements. We demonstrate this framework on a highly non-linear fluid-structure interaction problem: an underactuated, two-joint dogfish swimmer tasked with spatial target reaching using only joint angular accelerations. Starting from a propulsive seed policy that exhibits a one-sided steering bias, the agent autonomously discovers and refines a unified controller that robustly captures all canonical targets. Remarkably, without any retraining or target-specific branching, the synthesized control policy generalizes to unseen static targets and dynamically curved pursuit trajectories. The auditable evolve log reveals an emergent control architecture built upon traveling-wave propulsion, body-frame target guidance, yaw-rate feedback, signed mean-tail curvature, and adaptive cadence relief. Our results show that an autonomous scientific agent can successfully transform accumulated physical evidence into robust, mathematically readable control policy, while maintaining a fully traceable process of scientific discovery.

Human manipulation videos are a convenient and intuitive source for robot learning. However, directly transferring human dexterity to robots remains challenging due to perception errors and embodiment gap. To address this, we introduce Video2Sim2Real, a full-stack framework for autonomous skill acquisition from a single human manipulation video. Our framework first uses off-the-shelf foundation models to reconstruct a simulator-ready digital twin and extract robot and object motion priors. Rather than treating the extracted robot motion as a reliable reference throughout execution, our key idea is to recover and leverage the most fundamental sources of supervision from the demonstrated skill: We identify object-centric keyframes to optimize the corresponding robot configurations using object information from the simulator, and use these configurations as anchors that refine the robot motion such that it ultimately has the desired impact on the environment. To bridge the remaining sim-to-real gap, we introduce a sim-to-real strategy that decouples robustness to noisy and incomplete perception from variations in hand-object interaction dynamics. Specifically, we learn to recalibrate robot configurations from noisy real-world point clouds via IL, and leverage residual RL to perform local finger-level adaptations to ensure for robust and effective interactions. Finally, a collision-aware motion planning module enables spatial generalization to novel object configurations. Across several everyday manipulation tasks, Video2Sim2Real improves simulated task success, safety, and trajectory coherence over numerous baselines, and achieves better sim-to-real transfer than existing techniques. These results demonstrate a promising path toward autonomous dexterous skill acquisition from human videos.

Large language model (LLM) agents are increasingly deployed in long-running settings where improving through experience at test time becomes important. A common approach is to update an explicit memory after each interaction to guide future decisions. However, most existing methods rely on hand-designed prompting rules, making it difficult to align memory updates with downstream objectives over multi-step horizons consistently. We propose MemoPilot, a plug-in memory copilot that explicitly trains the memory update process to improve a frozen LLM's performance across sequential interactions. We formulate memory updating as a multi-turn decision problem and optimize it end-to-end with multi-turn GRPO. Our training recipe introduces (i) a turn-wise reward signal and (ii) a context-independent, turn-level advantage estimation across rollouts, enabling finer-grained credit assignment and more stable training in multi-turn settings. We evaluate MemoPilot on two testbeds: multi-round Rock-Paper-Scissors (RPS) and Limit Texas Hold'em (LHE). Across both environments, MemoPilot substantially improves test-time learning of a frozen player over strong baselines, ranking first in Elo ratings on both games (1762 on LHE and 1590 on RPS) and outperforming all baseline memory methods and proprietary models, including DeepSeek-V3.2.

A scientific hypothesis is the first step in research and undergoes experimental validation, yet it also reflects a deep understanding of and reasoning about scientific phenomena. We introduce DN-Hypo-Pipeline, an AI-powered workflow based on large language models, designed to support structured scientific thinking and hypothesis generation by leveraging scientific explanations as prior knowledge. This pipeline assists researchers in deriving novel hypotheses from existing literature. Given the explanandum (i.e., the conclusion) of a research paper, it identifies underlying laws, theories, and principles, and reconstructs a new, yet-to-be-verified explanation for the observed phenomenon. We evaluated DN-Hypo-Pipeline in the field of data science modeling using three highly cited papers. Statistical inference, supported by both LLM-as-judge assessment and human expert evaluation, demonstrates that our pipeline is more effective than direct generation methods. Additionally, we validated the two highest-scoring generated hypotheses by developing corresponding novel algorithms, which outperformed the baseline models presented in the original papers. Beyond application in data science, DN-Hypo-Pipeline provides a theoretical framework that not only encompasses theory-guided data science modeling methods but also reveals a more fundamental structure of the modeling process. Moreover, this approach is essentially a generalization of theory-guided modeling, offering potential for extension to other domains and across a broader range of scientific disciplines.

Agent skills provide a lightweight mechanism for extending general-purpose agents, but their open format exposes them to skill-poisoning attacks. A practically dangerous injection must stay invisible: if executing the payload derails the user's legitimate task, the resulting failure signal invites inspection of the skill. We therefore evaluate attacks by Attack Success Rate, which requires the injected payload to execute and the user's task to still pass its verifier in the same trial. Prior skill-poisoning attacks face a reliability-stealth trade-off under this lens: YAML-header injections are reliably loaded but easily inspected, whereas stealthier body injections that place explicit malicious commands in the skill prose are less reliable because out-of-context commands invite the agent's own suspicion. We introduce POISE, a position-aware attack that compresses the trigger into a single, benign-looking body instruction, placing it at a feasible position and using a context-aware generator to blend it with nearby setup or prerequisite steps. On Skill-Inject with codex+gpt-5.2, POISE achieves an 89.3% ASR, 28.0 points above a random-placement body baseline and 2.6 points above a YAML-only baseline, while retaining the stealth advantage of body placement. That stealth is the decisive margin: because legitimate skill bodies naturally require privileged tool operations, LLM scanners are hyper-sensitive, falsely flagging 74.6% of clean skills on average across four judges and both benchmarks. Blending into these false alarms, POISE causes only 5.6% of poisoned variants to gain a new high-risk alert over their clean baselines, rendering current static defenses ineffective.

Current image editing software often hinges on fixed filters or expert tuning, leaving a gap between amateur users' intent and outcomes. Creations by generative models may contain artifacts, implausible details, or stylistic drift away from photorealism and offer little insight into why an edit was made. We propose IEA, a conversational Image Editing Agent that learns to operate parameterized tools in an explicit, interpretable action space. IEA is trained via a three-stage multitask pipeline: (1) SFT on distilled expert edits, (2) GRPO with rewards for likeness improvement, tool usefulness, and intent summarization, and (3) large-scale synthetic fine-tuning to jointly master image editing, refinement, and user intent summarization. By manipulating 16 editing tools step by step, IEA produces transparent edit traces that can be inspected and debugged. In quantitative experiments, it attains a lower pixel distance on the edit task and a higher ROUGE-L on the summary task than strong baselines. In user studies, it ranks best among tool-calling methods for instruction following while surpassing generative methods in overall perceptual quality. Our results validate interpretable, tool-centric VLMs as a reliable path to human instruction-guided image retouching.

Repeated reference games test whether interlocutors replace their initially long descriptions with shorter, partner-specific conventions grounded in shared interaction history. Prior work shows that multimodal LLMs fail to become more efficient across rounds, although they align on the labels they use. How can we determine whether this alignment reflects partner-specific grounding rather than a shared task vocabulary? We address this question by comparing capable multimodal agent dyads with human dyads from the KTH Tangrams corpus. Our novel methodological contribution is a constrained pseudo-dyad baseline that matches the original referential task structure, but breaks partner history. This baseline enables us to test whether the observed label alignment depends on interaction with a specific partner. Across three analytic layers (task competence, description strategy, alignment dynamics), we find clear differences. Humans reduce effort through entrainment, compressing descriptions and increasing label alignment with partners. Agents instead maintain fixed effort levels, producing verbose descriptions from round one, with near-ceiling label overlap that is statistically indistinguishable between real and pseudo dyads. MLLMs thus achieve coordination without convention, succeeding by verbose description rather than by forming the compact, history-dependent referring expressions characteristic of human dialogue.

When an LLM agent fails — issues a refund it should not have, calls the wrong tool, leaks data — existing tooling answers what happened (observability) or whether it passed (evaluation), but not which step caused the failure. The obvious heuristics are wrong: the step that executes the harmful action is usually not the step that decided on it, and LLM-judge attribution is correlational and unreliable (state-of-the-art step-level accuracy on the Who&When benchmark is about 14%). We present Causal Agent Replay (CAR), which answers the question by intervention: it models an agent run as a structural causal model, applies a do-operation to a step, and re-executes the trajectory forward under the same stochastic policy, measuring the shift in the outcome distribution. We define an intervention algebra over agent steps, a single-step contrastive estimator whose point-of-commitment rule resolves a confound specific to stochastic run-forward, and a budget-bounded Monte-Carlo Shapley estimator that splits credit across interacting steps. Every effect is reported with confidence intervals. We validate against synthetic structural causal models with planted ground truth: the contrastive estimator recovers the pivotal step, and Shapley recovers a two-step interaction (0.44, 0.45, ~0; efficiency sum 0.909 versus the analytic 0.91). CAR is open source and runs on hosted or free local models.

LLM-based scientific agents have shown strong capacity for autonomous research, yet their safety layers remain structurally divorced from core reasoning: they inspect pipeline outputs rather than shaping the deliberation that produces them. This separation opens two failure modes: safety signals accumulated at one stage are discarded before the next, and sequences of individually benign tool calls can compose into harmful outcomes that no single-step filter detects. To address these challenges, we introduce SciTrace, a framework that weaves safety reasoning into every stage of the scientific agent pipeline. SciTrace couples two complementary mechanisms: a Safety-Intrinsic Reasoning Loop (SIR) that maintains a cumulative risk state across the Thinker, Experimenter, Writer, and Reviewer stages through joint task-and-safety deliberation, and a Compositional Tool-Chain Verifier (CTV) that performs trajectory-aware safety checks before execution, catching risks that surface only across multi-step tool sequences. Evaluated on 240 high-risk research tasks and 120 tool-related risk tasks spanning six scientific domains, SciTrace achieves state-of-the-art (SOTA) safety among compared frameworks across four backbone models: it consistently improves tool call safety and adversarial robustness while preserving scientific output quality, and it uncovers 78.8% of the compositional tool-chain escapes that single-step monitors miss. The project website is available at https://opensciagent.github.io/SciTrace/.

Multi-agent large language model (LLM) systems are rapidly emerging, yet transparency, a cornerstone of responsible AI, remains under-defined in these distributed architectures, which have complexities of inter-agent coordination and orchestration. In this paper, we present one of the first empirical study of how early adopters of multi-agent LLM systems, who are both the builders and users, understand and practice transparency. We conducted semi-structured interviews with 13 early adopters in [Large Technology Organization] and applied thematic analysis to identify recurring patterns. Participants articulated divergent yet complementary framings of transparency, including reproducibility, debugging, boundary-setting, visualization, and auditing. These perspectives spanned questions of what transparency entails, why it matters, and how it is achieved. We synthesize these into a multidimensional framework, which is developer, user, and governance-focused positioning transparency as a situated socio-technical practice that informs future HCI and AI design and research around aligning expectations and capacities of their intended audiences.