Agent 进展

Large language model agents now act on codebases, browsers, operating systems, calendars, files, and tool ecosystems, but the benchmarks used to evaluate them are fragmented: each emphasizes a different unit of measurement (final task success, tool-call validity, repeated-pass consistency, trajectory safety, or attack robustness). A line of 2024-2025 work has converged on the diagnosis that a single accuracy column is no longer the right unit of comparison for deployable agents. AgentAtlas extends this line of work with four components: (i) a six-state control-decision taxonomy (Act / Ask / Refuse / Stop / Confirm / Recover); (ii) a nine-category trajectory-failure taxonomy with two orthogonal hierarchical labels (primary_error_source, impact); (iii) a taxonomy-aware vs. taxonomy-blind methodology that measures how much of a model's apparent capability comes from the supervision in the prompt; and (iv) a benchmark-coverage audit mapping fifteen agent benchmarks against six behavioral axes. To demonstrate the methodology we run a small fixed eight-model set (1,342 generated items, four frontier closed and four open-weight) under both prompt modes. Removing the explicit label menu drops every model's trajectory accuracy by 14-40 pp to a tight 0.54-0.62 floor regardless of family, and no single model wins on all three of control accuracy, trajectory diagnosis, and tool-context utility retention. We treat the synthetic run as a measurement-protocol demonstration, not a benchmark release.

Large language models (LLMs) have enabled web agents that follow natural language goals through multi-step browser interactions. However, agents fine-tuned on specific trajectories and domain often struggle to generalize out of domain, and offline training can be compute-inefficient due to noisy, redundant trajectories and long accessibility-tree (AXTree) states. To address both issues, we propose Weasel, a trajectory selection method for offline training of web agents. Weasel selects a fixed-budget subset of trajectory steps by optimizing an objective that balances unary importance with pairwise diversity over states, websites, and interaction patterns, solving efficiently with a greedy algorithm. We further improve efficiency with target-centered AXTree pruning that keeps only content around the ground-truth action target, and we mitigate style mismatch for reasoning-native models by replacing expert traces with model-generated, style-consistent rationales. Across AgentTrek and NNetNav training datasets, evaluations in WebArena, WorkArena, and MiniWob, and experiments with Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel improves out-of-domain performance while reducing training cost, producing roughly 9.7-12.5$\times$ training speedups over standard fine-tuning. We make the code available at https://github.com/fatemehpesaran310/weasel.

Vision-language models (VLMs) are used as high-level planners for embodied agents, translating natural language instructions and visual observations into action plans. While prior work has studied abstention in LLMs, existing benchmarks are largely text-only and do not capture the perceptual grounding and physical constraints inherent to embodied robotics environments. In such settings, abstention requires recognizing when instructions are ambiguous, physically infeasible, based on false premises, or otherwise unresolvable given the available sensory modalities and context. To address this gap, we introduce a taxonomy to categorize abstention in the context of embodied robotics and present RoboAbstention, a scalable and auditable framework for generating abstention instructions grounded in images gathered from five robotics datasets. RoboAbstention instantiates the taxonomy through a three-phase pipeline: (1) structured visual grounding, (2) deterministic constraint derivation, and (3) controlled instruction generation via category-specific templates. This enables the construction of a diverse dataset with verifiable abstention conditions. We evaluate several frontier VLMs and find that all models exhibit significant weaknesses in abstention, including those with advanced reasoning capabilities. The best-performing model, Gemini 2.5 Flash, abstains on only 39.0% of our 6,069 benchmark instructions, while the embodied planner Gemini Robotics ER 1.6 Preview abstains on just 16.5%. We further explore methods for improving abstention in VLM planners, such as defensive prompting and in-context learning, and find that these interventions substantially improve performance, reaching 93.6% abstention rate for Gemini Robotics ER 1.6 Preview and 88.6% for GPT 5.4 Mini, yet no approach fully solves the problem. We open-source RoboAbstention at https://purseclab.github.io/RoboAbstention/.

We introduce WildRoadBench, a wild aerial road-damage grounding benchmark that couples direct visual grounding by vision-language models with autonomous research-and-engineering by LLM-driven agents on a single professionally annotated UAV corpus. The same image set and the same per-class AP_50 metric are evaluated under two protocols. The VLM Track measures whether a fixed VLM can localise domain-specific damage from one image and one short prompt under a unified prompting, decoding and parsing pipeline. The Agent Track measures whether an autonomous agent, given only a written task brief, a small exploratory slice and a fixed interaction budget, can search the public web, adapt pretrained components, write training and inference code, and submit predictions through a scalar-feedback oracle on a hidden holdout. We benchmark a broad pool of closed-source frontier models and open-source VLMs together with several frontier LLM-driven agents. Both routes remain far from reliable performance in this wild setting: closed-source frontier models lead the VLM leaderboard but still leave more than half of the metric on the table; open-source grounders plateau well below them, and newer generations or reasoning-style variants do not consistently improve grounding; small targets collapse for every open-source model; agents lag the strongest VLM despite richer affordances, and several fail to land a valid submission within the budget. We release the code and data at https://anonymous.4open.science/r/wildroadbench-0607 to support reproducible follow-up research.

Training large multimodal models (LMMs) via reinforcement learning (RL) to natively invoke video-processing tools (e.g., cropping) has become a promising route to long-video understanding. However, existing native-RL methods dispatch tool calls sequentially (i.e., one per turn): a single wrong crop propagates errors without peer correction, multi-turn tool calls corrupt context, and inference cost scales linearly with the number of turns. We introduce ParaVT, the first multi-agent end-to-end RL-trained framework for Parallel Video Tool calling, dispatching multiple time-window crops in a single turn for cleaner context and better fault tolerance. Yet applying standard RL to ParaVT reveals an obstacle we term the Tool Prior Paradox: the pretrained tool priors that enable tool exploration also destabilize cold-started structural format and expose the skip-tool reward shortcut under temperature sampling. A cross-model contrast on a weaker-prior LMM supports this claim: format stays stable but RL elicits zero tool calls, indicating that prior strength is the shared driver of both format collapse and tool exploration. We propose PARA-GRPO (Parseability-Anchored and Ratio-gAted GRPO), which augments standard RL with two complementary mechanisms: (i) a targeted format reward applied only at the structural-token positions most prone to collapse, and (ii) a per-prompt frame-budget randomization that creates training prompts where calling the tool yields a measurable reward signal over skipping it. Across six long-video understanding benchmarks, ParaVT improves over the Qwen3-VL baseline by +7.9% on average, with PARA-GRPO lifting training-time format compliance from 0.13 to 0.64. As tool capabilities become increasingly internalized in modern LMMs, RL must cooperate with the resulting priors, and ParaVT offers a general recipe for agentic RL. Code, data, and model weights are publicly available.

Recent advances in Vision-Language Models (VLMs) facilitate a new class of embodied AI systems, where these models are integrated into physical platforms, e.g. robots and autonomous vehicles, to interpret visual scenes and execute natural language commands in diverse environments. Previous research has introduced jailbreak attacks and defenses for embodied AI. Their evaluations, however, rely on ad-hoc datasets, limited metrics, and emphasize attack success while neglecting the trade-off between security and the ability to follow benign commands. Existing benchmarks and evaluation frameworks either target traditional chat-based models or focus on non-adversarial safety evaluation for embodied AI; neither captures the adversarial risks, inputs, consequences, and evaluation criteria necessary for jailbreak attacks in embodied AI systems. In this paper, we address this gap with RoboJailBench, which consists of three core components. We establish a security taxonomy derived from ISO standards, regulatory rules, and documented incidents. This effort yields 18 categories of security violation consequences for embodied AI. We introduce an intent contrast dataset pipeline that augments existing datasets with paired adversarial and benign goals to measure both security and utility. Lastly, we provide an evolving repository with standardized metrics and a unified process for assessing and integrating new attacks and defenses. With this benchmark, we construct a new taxonomy-balanced dataset and augment five existing datasets. We integrate four attacks and two defenses to evaluate their performance on leading embodied VLMs. This benchmark provides the first standardized evaluation framework for jailbreak attacks in embodied AI and supports future research. We release our code, datasets, and artifacts, and maintain a leaderboard at https://purseclab.github.io/benchmark-for-robotics-security.

LLM discovery and optimization systems are increasingly applied across domains, implementing a common propose-evaluate-revise loop. Such optimization or discovery progresses via context conditioning on received feedback from an environment. However, as modern LLM agents are increasingly complex in their structure, it is difficult to evaluate which components contribute the most, and when and how this exploration may fail. We answer these questions through three controlled experiments. Our findings: (1) In pure black-box optimization, LLMs act as greedy optimizers. (2) In zero-shot kernel generation, providing explicit input-size information has no measurable effect, models converge to the same kernel parameters regardless of size or temperature, as though the size instruction were invisible. Moreover, when tasked to perform kernel optimization for uncommon kernel sizes, performance sharply degrades regardless of the language used. (3) In feedback-loop kernel optimization, CUDA improves monotonically under iterative feedback, while TVM IR actively degrades, which demonstrates that kernel optimization degrades when models operate with low-density language. Our results conclude that LLMs in code optimization tasks highly depend on pretrained priors rather than provided feedback or agentic structure.

Do stock safety-aligned language models and their uncensored or abliterated derivatives behave differently when run as autonomous security agents? Single-turn refusal benchmarks cannot answer this question: security agents must inspect repositories, call tools, and produce vulnerability evidence inside authorized sandboxes. We present a trace-based benchmark of 30 local vulnerability-analysis tasks with fixed tools, deterministic success predicates, redaction rules, and grounding checks, and compare four stock models against uncensored or abliterated derivatives: Gemma 4 31B, Gemma 4 26B A4B, Qwen2.5-Coder 7B, and Llama 3.1 8B. The artifact contains 1,500 security-agent traces and 800 non-security control traces. The Gemma pairs show large less-restricted gains on security tasks: 14.0% versus 0.7% success for 31B and 10.7% versus 0.0% for 26B, with higher mean grounding (3.91 versus 3.27 and 4.12 versus 1.64 out of five) and 0.0% refusal, suppressed-action, and unsafe-action rates in the 31B traces. However, controls and non-Gemma pairs rule out a clean security-specific or universal less-restricted effect: Gemma gaps also appear on ordinary coding tasks, Qwen2.5-Coder success is lower for the less-restricted derivative (2.0% versus 5.3%), and the abliterated Llama derivative fails the tool protocol. Across all families, hard proof-of-trigger and patch-verification tasks remain unsolved. These results show that safety alignment effects in autonomous security agents should be measured at the system level, separating refusal, unsafe action, tool reliability, and evidence grounding rather than treating refusal rate as the safety signal.

Large Multimodal Models (LMMs) have recently emerged as promising backbones for GUI-agent models, where high-resolution GUI screenshots are introduced to the prompts at each iteration step. However, these screenshots exhibit highly non-uniform spatial information density: large regions may carry little information and are visually homogeneous, while key text and icons may require high visual fidelity. Existing approaches to this problem either require additional training or rely on attention-based token compression, ignoring the structured layout and spatial redundancy of GUI screenshots. To fill the gap, this paper proposes AquaUI, a training-free inference-time token reduction method for GUI agent models that utilizes the non-uniform information density in screenshots. AQuaUI constructs an adaptive quadtree on each screenshot input and keeps one representative merged token per leaf of the quadtree. AQuaUI preserves the spatial positions of retained tokens throughout the pipeline to ensure that all position-encoding stages remain consistent. To further improve temporal consistency across multi-step GUI interactions, we propose a conditional quadtree algorithm that leverages the continuity between consecutive screenshots within a single request. Specifically, it refines the current quadtree using previous quadtrees as references, helping preserve fine-grained regions across static or mildly shifted GUI states. We implement AQuaUI on state-of-the-art GUI agent models and conduct experiments on standard grounding and navigational benchmarks. AQuaUI consistently shows improved accuracy-efficiency trade-offs over prior baselines. Notably, on GUI-Owl-1.5-32B-Instruct, AQuaUI achieves up to 13.22% speedup and 29.52% fewer visual tokens while retaining 99.06% of full-token performance, suggesting that the spatial redundancy of GUI screenshots can be exploited at inference without retraining.

LLM agents organize behavior through skills - structured natural-language specifications governing how an agent reasons, retrieves, and responds. Unlike monolithic prompts, skills are multi-field artifacts subject to hard platform constraints: description fields are truncated for routing, instruction bodies are compacted via progressive disclosure, and co-resident skills compete for limited context windows. These constraints make skill optimization inherently multi-objective: a skill must simultaneously maximize task performance and satisfy platform limits. Yet existing prompt optimizers either ignore these trade-offs or collapse them into a weighted sum, missing Pareto-optimal variants in non-convex objective regions. We introduce MOCHA (Multi-Objective Chebyshev Annealing), which replaces single-objective selection with Chebyshev scalarization - covering the full Pareto front, including non-convex regions - combined with exponential annealing that transitions from exploration to exploitation. In our experiments across six diverse agent skills - where all methods share the same multi-objective mutation operator and baselines receive identical per-objective textual feedback - existing optimizers fail to improve the seed skill on 4 of 6 tasks: 1000 rollouts yield zero progress. MOCHA breaks through on every task, achieving 7.5% relative improvement in mean correctness over the strongest baseline (up to 14.9% on FEVER and 10.4% on TheoremQA) while discovering twice as many more Pareto-optimal skill variants.

Reinforcement learning can train LLM agents from sparse task rewards, but long-horizon credit assignment remains challenging: a single success-or-failure signal must be distributed across many actions. Existing methods rely on trajectory-level rewards or proxy signals, without fully leveraging per-step environmental feedback. Multi-turn agent settings are underexplored, where feedback can include error messages, page changes, observations, or reference trajectories. We systematically study five feedback sources and two insertion granularities and introduce SERL, a selective environment-reweighted learning framework. SERL uses the task reward to determine update direction, while environment feedback adjusts placement and magnitude, focusing on critical actions. On ALFWorld and WebShop, SERL achieves 90.0% and 80.1% success, outperforming strong RL and distillation baselines. Analysis shows that grounded, action-relevant feedback at meaningful points consistently outperforms indiscriminate use of longer or richer context.

Agent Skills, structured packages of procedural knowledge loaded into an LLM agent at inference time, are widely reported to improve task pass rates by an average of 16.2~percentage points across diverse domains. Yet the same benchmarks show wide variance, with 16 of 84 tasks suffering negative deltas when Skills are introduced. The community has not yet articulated a clean mechanism for when Skills help and when they are merely redundant overhead. We re-analyze a recently published 180-run controlled study of an MCP-grounded autonomous Capture-the-Flag (CTF) agent under four documentation conditions of increasing richness (55, 1{,}478, 1{,}976, and 4{,}147 lines), and show that these conditions correspond almost exactly to a No-Skills, Experiential-Skills, Curated-Skills, and Comprehensive-Skills ablation. In offensive cybersecurity, a domain not deeply covered by existing Skills benchmarks, the marginal benefit of Skills collapses. The spread between the no-Skills and full-Skills conditions is only 8.9~pp ($p = 0.71$, $χ^2$; $p = 0.25$, Cochran--Armitage trend test; five of six pairwise Cohen's $h$ values fall below the $0.2$ small-effect threshold). We argue that the missing variable is environment-feedback bandwidth. When an agent's tool layer returns strict, schema-validated, low-latency observations, the environment itself supplies the procedural correction signal that Skills are normally needed to provide. As a result, the marginal benefit of curated Skills diminishes substantially, and, in some cases (e.g., our timing side-channel setting), actively degrades performance. We articulate a falsifiable hypothesis, sketch its design implications for compound AI systems, and will release the reanalysis pipeline to support replication.

Recent work pairs LLMs with evolutionary search to iteratively generate, modify, and select code using task-specific feedback. These systems have produced strong results in mathematical discovery and algorithm design, yet a fundamental question remains: what do they actually evolve? Progress is typically summarized by the best score a run reaches under a task-specific evaluator, but that score can reflect several different mechanisms: new algorithmic structure, re-tuning an existing strategy, recombining ideas already in the model's internal knowledge, or overfitting to the evaluator. Distinguishing these mechanisms requires inspecting the search process itself, not only its final outcome. We introduce EvoTrace, a dataset of evolutionary coding traces spanning four evolutionary frameworks, reasoning and non-reasoning models, and 16 tasks across mathematics and algorithm design. To analyze these traces, we develop EvoReplay, a replay-based methodology that reconstructs the local search states behind high-scoring solutions and tests controlled interventions, including adjusting constants, removing program components and substituting models or prompting contexts. We annotate every code edit in EvoTrace with one of nine recurring edit types using an LLM-as-judge pipeline validated against blind human re-annotation. Across EvoTrace, most score gains come from a small subset of these edit types. We further find a deterministic cycling pattern: about 30% of code lines added during search are byte-identical re-introductions of previously-deleted lines, present throughout nearly every run. These results show that benchmark gains in evolutionary coding agents can arise from qualitatively different mechanisms, only some of which correspond to new algorithmic structure. EvoTrace enables more diagnostic evaluation of evolutionary coding agents beyond final benchmark scores.

Production LLM agents combine stochastic model outputs with deterministic software systems, yet the boundary between the two is rarely treated as a first-class architectural object. This paper names that boundary the stochastic-deterministic boundary (SDB): a four-part contract among a proposer, verifier, commit step, and reject signal that specifies how an LLM output becomes a system action. We argue that the SDB is the load-bearing primitive of production agent runtimes. Around this primitive, we organize agent runtime design into three concerns: Coordination, State, and Control. We present a catalog of six runtime patterns that compose the SDB differently across conversational, autonomous, and long-horizon agents: hierarchical delegation, scatter-gather plus saga, event-driven sequencing, shared state machine, supervisor plus gate, and human in the loop. For each pattern, we trace its lineage to distributed-systems concepts and identify what changes when the worker is stochastic. The paper contributes a five-step methodology for selecting runtime patterns, a diagnostic procedure that maps production failures to pattern weaknesses, and a failure mode called replay divergence, in which LLM-based consumers of a deterministic event log produce different downstream outputs under model-version or prompt changes. A stylized reliability decomposition separates per-call model variance from architectural momentum, motivating the claim that as model variance decreases, pattern choice and SDB strength become increasingly important levers for long-run reliability. We apply the methodology to five workloads and provide one runnable reference implementation for a 90-day contract-renewal agent.

Modern software systems evolve rapidly under CI/CD practices, where tests are critical for quality. However, substantial code changes often render existing test cases obsolete, causing pipeline disruptions, reduced productivity, and compromised quality. Recent automatic test update approaches leverage LLMs to refine test cases via execution feedback and exact-matching context retrieval, prioritizing executability and line coverage but suffering three limitations: (1) neglecting test assertion adequacy, weakening fault detection; (2) relying on coarse line coverage instead of specific uncovered lines/branches; (3) using exact-matching retrieval, which fails for LLM hallucinated queries. To address these, we propose MuMuTestUp, a mutation-guided multi-agent framework with three specialized agents: Mutation Analysis (strengthens assertions via surviving mutants), Coverage Analysis (generates targeted repair instructions for uncovered lines/branches), and Semantic Retrieval (handles hallucinations via semantic-similarity search). We also construct PRBENCH, a 571-sample pull-request-level dataset from 10 open-source Java projects (validated for cross-commit update scenarios). Evaluations against state-of-the-art baselines use both open-source (Deepseek-V3.2) and closed-source (GPT-4.1) LLMs.

A growing body of work explores how Large Language Models (LLMs) can be embedded in trading systems as agents that perceive market information, retrieve context, reason about decisions, emit tradable actions, and adapt under market feedback. This paper reframes LLM-based trading agents as expert-system decision pipelines and presents an audit-oriented evidence map of 77 included studies in a protocol-coded snapshot screened through 2026-03-09. A primary empirical subset (n=19) satisfies the minimum boundary of Action Output plus Closed-Loop Evaluation; the remaining 58 included studies are retained as background and design context. The central empirical finding is protocol incomparability: within the primary subset, only 2/19 studies report extractable time-consistent split protocols, 1/19 reports an explicit transaction-cost model, 1/19 documents universe or survivorship handling, 11/19 report execution timing or semantics, 15/19 are coded as R0, and no study reaches R3 reproducibility. We therefore use Architecture-Capability-Adaptation as a working analytical lens rather than a validated taxonomy, and we foreground the evidence ledger, reproducibility audit, and reporting checklist as the main contributions. The resulting survey shows that architectural experimentation is expanding rapidly, while comparable evaluation protocols, execution semantics, and reproducible artifacts remain the field's immediate bottlenecks.

Explicit software architecture models are essential artifacts for communicating, analyzing, and evolving complex software-intensive systems. In ROS~2-based robotic systems, however, structural (de-)composition and integration semantics are often only implicitly encoded across distributed artifacts such as source code and launch files, making recovery of hierarchical architecture particularly difficult. Existing approaches mainly focus on node-level entities and communication wiring, while providing limited support for recovering hierarchical structural (de-)composition across multiple abstraction levels. In this paper, we extend our previously proposed blueprint-guided LLM-assisted architecture recovery pipeline for ROS~2 systems through two major enhancements: (1) refined prompting to improve the consistency and controllability of architecture synthesis, and (2) a staged recovery strategy based on multi-level intermediate architectural representations that incorporate the atomic ROS node list and launch file dependencies, thereby enabling structurally constrained reconstruction across multiple abstraction levels. The approach is evaluated on a real-world automated product disassembly system based on cooperative robotic arms and heterogeneous ROS~2 artifacts. Compared to our previous work, the considered case study exhibits substantially higher integration complexity and richer functionality. The results demonstrate improved structural consistency, scalability, and robustness of architecture recovery, while also revealing remaining challenges related to dynamic integration semantics in large-scale ROS~2 systems.

Long-horizon embodied agents increasingly delegate navigation, search, approach, and manipulation to specialist executors. As these executors become stronger, the main bottleneck shifts from local skill execution to maintaining a coherent task frontier across planning, monitoring, memory, and execution. We study task-state misalignment, a task-level consistency failure in which the planner's active stage, runtime evidence, remembered context, and delegated executor no longer justify the same next-step decision. This failure can lead to unsupported handoffs, stage lock, executor-context mismatch, and unnecessary replanning. We propose ContextFlow, an inspectable alignment framework that represents stages as explicit contracts, converts runtime observations into evidence packets, and applies scoped updates including continue, refine, transfer, promote, and repair. ContextFlow keeps specialist executors responsible for local closed-loop control while making task-frontier alignment explicit and auditable. Experiments and demonstration traces on long-horizon embodied tasks illustrate how evidence-grounded scoped updates diagnose and mitigate recurring task-state failures.

While GUI agents have made significant progress in web navigation and basic operating system tasks, their capabilities in professional creative workflows remain largely underexplored. To bridge this gap, we introduce Cutverse, a benchmark designed to systematically evaluate autonomous GUI agents in realistic media post-production environments. We curate expert demonstrations across 7 professional applications (e.g., Premiere Pro, Photoshop), covering 186 complex, long-horizon tasks grounded in authentic editing workflows, involving dense multimodal interfaces and tightly coupled interaction sequences. To support scalable evaluation, we develop a lightweight parser that transforms raw screen recordings and low-level interaction logs into structured, compositional GUI action trajectories with precise grounding. Extensive evaluations reveal that existing agents achieve only 36.0% task success on realistic media editing tasks, underscoring the challenges posed by complex, long-horizon media post-production workflows in our benchmark.While current models demonstrate promising spatial grounding, multimodal alignment, and coordinated action execution, they remain limited in long-horizon reliability and domain-specific planning.

3D localization in Multimodal Large Language Models (MLLMs), including 3D object detection and 3D visual grounding, is fundamentally limited by camera intrinsic ambiguity: the same image admits different 3D scenes under different cameras. Existing MLLMs either ignore camera parameters and overfit to a canonical training intrinsic, or retrieve depth and 3D cues from external tools but treat the returned values as reference cues (numerical hints that the model is free to interpret implicitly), both preventing camera information from being deterministically propagated into the prediction. We propose an equation-anchored tool-use framework that re-purposes spatial tools as formula variables. The proposed framework proactively retrieves camera intrinsics and samples multi-point metric depths, writes the pinhole back-projection equation $\hat{X} = (u_c - c_x)\bar{Z}/f_x$ explicitly in Chain-of-Thought (CoT), and substitutes tool outputs into the formula before regressing the final 9-DoF bounding box. On both 3D object detection and 3D visual grounding tasks under rescaled camera intrinsics from $0.5\times$ to $1.5\times$, our method outperforms RGB-only and tool-augmented baselines, with significant gains where the camera deviates most from the training scale. Code and data will be released.

Recent studies in difficulty-controlled reading comprehension item generation have leveraged large language models (LLMs) to produce items by adjusting difficulty-related features. However, existing methods typically rely on a single-agent prompting approach, which often fails to consistently satisfy specified feature constraints, resulting in items that deviate from the target difficulty level. To address this limitation, we introduce MAFIG, a Multi-agent Framework for Feature-constrained Item Generation, where multiple LLM agents and feature-specific evaluators collaborate to generate and iteratively revise items based on intended constraints. Furthermore, to verify the efficacy of MAFIG in difficulty control, we propose a method for constructing a sequence of feature constraint sets that yield items with monotonically increasing difficulty. Experimental results demonstrate that MAFIG generates items that adhere to target constraints at a significantly higher rate than baselines, achieving robust difficulty control through the difficulty-calibrated constraint sequence.

As Socially Interactive Agents (SIAs) become increasingly integrated into daily life, the ability to calibrate user trust to an agent's actual capabilities would help ensure appropriate usage of these agents. In this paper, we explore the capacity of Large Language Models (LLMs) to generate multimodal behaviors (verbal, vocal, gestural, and facial expression modalities) that reflect varying levels of ability and benevolence, two key dimensions of trustworthiness. We propose a novel method for automatically generating behaviors aligned with specific levels of these traits, a first step towards enabling nuanced and trust-calibrated interactions. By analyzing a large dataset of multimodal transcripts generated by LLMs, we demonstrate that GPT-5.4 is able to produce coherent behavior across different modalities (text, intonation, facial expression, and gesture). Using Random Forest feature importance analysis, we show that the generated behaviors align with theoretical expectations for ability and benevolence. However, we also find that when gender is specified in the prompt, LLMs tend to reproduce societal gender stereotypes, associating male agents' behaviors with high ability and female agents' behaviors with high benevolence. To validate our approach, we conducted a user study on Prolific using a within-subjects design. Participants perceived different levels of ability and benevolence in the generated behaviors align with the intended instructions.

Segmentation has become easy when the concept is known, requiring retrieval of a learned visual grounding from text. It remains hard for open ad-hoc concepts, where the grounding may not exist as one learned mask and must often be constructed from image evidence through parts, relations, exclusions, and collections. We propose a Vision-guided Ad-hoc Segmentation Agent (VASA), the first vision harnessing agent for open ad-hoc segmentation. VASA is training-free and couples a VLM agent, a segmentation foundation model, and a visually grounded workflow. Rather than revising text prompts alone, VASA uses a persistent working mask to reason, construct, and validate a solution. It plans visual operations, invokes segmentation tools, inspects results, edits the mask, and recovers from errors. We construct PARS, a new benchmark that turns part-level labels in PartImageNet into open ad-hoc concepts through long-form definition queries. On PARS, VASA outperforms open-vocabulary, reasoning-based, and agentic baselines, surpassing SAM3 Agent by 14-25%. On RefCOCOm, a standard multi-granularity referring segmentation benchmark, VASA improves over SAM3 Agent by 5-9% and over other agentic baselines by up to 20%. These results validate agentic visual construction for open ad-hoc segmentation. Our work points to a path for AI agents beyond wrapping foundation models as tools: Programming them with task knowledge, VLM behavior, visual routines, working memory, and failure-aware workflows.

The inspection of electrical power line insulators is essential for ensuring grid reliability and preventing failures caused by damaged or degraded insulation components. In recent years, Unmanned Aerial Vehicles (UAVs) combined with deep learning-based vision systems have emerged as an effective solution for automating this process. However, insulator fault detection remains challenging due to small defect regions, heterogeneous fault patterns, complex backgrounds, and varying imaging conditions. To address these challenges, this paper proposes an optimized YOLO26-MoE, a novel object detection architecture that integrates a sparse Mixture-of-Experts (MoE) module into the high-resolution branch of the YOLO26 detector. The proposed modification enables adaptive feature refinement for subtle and diverse fault patterns while preserving the efficiency of a one-stage detection framework. Hyperparameter optimization, final training, and evaluation were coordinated through a tool-augmented Large Language Model (LLM) agent. The proposed model achieved 0.9900 mAP@0.5 and 0.9515 mAP@0.5:0.95, outperforming the latest YOLO versions. These results demonstrate that the proposed model provides an effective and reliable solution for UAV-based insulator fault detection.

We present OpenComputer, a verifier-grounded framework for constructing verifiable software worlds for computer-use agents. OpenComputer integrates four components: (1) app-specific state verifiers that expose structured inspection endpoints over real applications, (2) a self-evolving verification layer that improves verifier reliability using execution-grounded feedback, (3) a task-generation pipeline that synthesizes realistic and machine-checkable desktop tasks, and (4) an evaluation harness that records full trajectories and computes auditable partial-credit rewards. In its current form, OpenComputer covers 33 desktop applications and 1,000 finalized tasks spanning browsers, office tools, creative software, development environments, file managers, and communication applications. Experiments show that OpenComputer's hard-coded verifiers align more closely with human adjudication than LLM-as-judge evaluation, especially when success depends on fine-grained application state. Frontier agents struggle with end-to-end completion despite partial progress, and open-source models exhibit sharp drops from their OSWorld-Verified scores, exposing a persistent gap in robust computer automation.

Reinforcement learning from verifiable rewards (RLVR) is a promising paradigm for improving large language model (LLM) agents on long-horizon interactive tasks. However, in partially observable environments, incomplete observations cause agent beliefs to drift over time, while delayed rewards obscure the causal impact of intermediate decisions, exacerbating temporal credit assignment challenges. To address this, we propose ReBel (Reward Belief), a process-level reinforcement learning algorithm that explicitly models structured belief states to summarize interaction history and guide subsequent policy learning. ReBel introduces belief-consistency supervision, converting discrepancies between predicted beliefs and observed feedback into dense self-supervised signals without requiring external step-wise annotations or verifiers. It also employs belief-aware grouping to compare trajectories under similar belief states, yielding more robust and lower-variance advantage estimates. We evaluate ReBel on challenging long-horizon benchmarks, including ALFWorld and WebShop. ReBel improves task success by up to $20.4$ percentage points over the episode-level baseline GRPO and increases sample efficiency by $2.1\times$. These results suggest that belief-aware self-supervision is a promising direction for reliable long-horizon decision-making under partial observability. Code is available at: https://github.com/Fateyetian/Rebel.git.

Large Language Model (LLM) agents increasingly act inside real workspaces, where tools and skills determine whether model reasoning becomes reliable action. Existing skills remain largely informal: Markdown skills and instruction packs encode procedures as long natural-language documents, while function calling, Model Context Protocol (MCP) servers, and framework tools structure individual actions but usually leave workflow state, policy enforcement, and completion discipline outside the skill itself. We introduce Formal Skill, a runtime-native abstraction that represents reusable capability with JSON metadata and action schemas, reliable Python executors, hook-governed control logic, Formal Skill routing, and skill-local runtime state. By moving reusable procedure from repeated prompt text into executable state machines and hook policies, Formal Skill gives agents a token-efficient and enforceable control surface. We implement the abstraction in FairyClaw, an open-source event-driven runtime for executable, observable, and composable Formal Skills. On Harness-Bench, FairyClaw obtains highly competitive average scores while using substantially fewer tokens, with especially strong results on tasks that expose the role of Formal Skill.

To enable reliable long-term interaction, LLM agents require a memory system that can faithfully store, efficiently retrieve, and deeply reason over accumulated dialogue history. Most existing methods adopt an extracted fact based paradigm: handcrafted static prompts compress raw dialogues into atomic facts, which are then stored, matched, and injected into downstream reasoning. Nevertheless, such fact-centric designs inevitably discard fine-grained details in original dialogues and fail to support deep reasoning over scattered isolated facts. Moreover, static prompts cannot maintain consistent extraction granularity across diverse dialogue styles. To address these limitations, we propose TriMem, which maintains three coexisting representation granularities, including raw dialogue segments anchored by source identifiers for storage fidelity, extracted atomic facts for efficient memory retrieval, synthesized profiles that aggregate dispersed facts into holistic semantic understanding for deep reasoning. We further adopt TextGrad-based prompt optimization, which iteratively refines extraction and profiling prompts via response quality feedback, achieving lifelong evolution without any parameter updating. Extensive experiments on LoCoMo and PerLTQA across multiple LLM backbones demonstrate that TriMem consistently outperforms strong memory baselines. The code is available at https://TMLR-TriMem.github.io .

Automatic generation of computer-aided design (CAD) models is a core technology for enabling intelligence in advanced manufacturing. Existing generation methods based on large language models (LLMs) often fall short when handling complex CAD models characterized by long operation sequences, diverse operation types, and strong geometric constraints, primarily because reasoning chains break and effective error-correction mechanisms are lacking. To address this problem, this paper proposes a memory-augmented reinforcement learning framework for CAD generation agents. The framework encapsulates the underlying geometric kernel into a structured toolchain callable by the agent and builds a closed-loop mechanism of design intent understanding, global planning, execution, and multi-dimensional verification. It also designs a dual-track memory module consisting of a case library and a skill library, and proposes a dynamic utility retrieval algorithm. By introducing reinforcement learning into retrieval and policy optimization, the agent can effectively avoid retrieval traps in which examples are semantically similar but geometrically infeasible, enabling online self-correction and continual evolution without additional large-scale annotated data. Experiments show that the proposed method significantly improves both the success rate and geometric consistency on complex CAD model generation tasks.

Large language model (LLM) agents increasingly operate over long and recurring external contexts, like document corpora and code repositories. Across invocations, existing approaches preserve either the agent's trajectory, passive access to raw material, or task-level strategies. None of them preserves what we argue is most needed for repeated same-context workloads: reusable orientation knowledge (e.g., what the context contains, how it is organized, and which entities, constants, and schemas have historically been useful) about the recurring context itself. We introduce PEEK, a system that caches and maintains this orientation knowledge as a context map: a small, constant-sized artifact in the agent's prompt that gives it a persistent peek into the external context. The map is maintained by a programmable cache policy with three modules: a Distiller that extracts transferable knowledge from inference-time signals, a Cartographer that translates it into structured edits, and a priority-based Evictor that enforces a fixed token budget. On long-context reasoning and information aggregation, PEEK improves over strong baselines by 6.3-34.0% while using 93-145 fewer iterations and incurring 1.7-5.8x lower cost than the state-of-the-art prompt-learning framework, ACE. On context learning, PEEK improves solving rate and rubric accuracy by 6.0-14.0% and 7.8-12.1%, respectively, at 1.4x lower cost than ACE. These gains generalize across LMs and agent architectures, including OpenAI Codex, a production-grade coding agent. Together, these results show that a context map helps long-context LLM agents interact with recurring external contexts more accurately and efficiently.