ICLR 2026 Papers

Agents are predominantly evaluated and optimized via task success metrics, which are coarse, rely on manual design from experts, and fail to reward intermediate emergent behaviors. We propose AutoLibra, a framework for agent evaluation, that transforms open-ended human feedback e.g. “If you find that the button is disabled, don’t click it again”, or “This agent has too much autonomy to decide what to do on its own” into metrics for evaluating fine-grained behaviors in agent trajectories. AutoLibra accomplishes this by grounding feedback to an agent’s behavior, clustering similar positive and negative behaviors, and creating concrete metrics with clear definitions and concrete examples, which can be used for prompting LLM-as-a-Judge as evaluators. We further propose two meta-metrics to evaluate the alignment of a set of (induced) metrics with open feedback: “coverage” and “redundancy”. Through optimizing these meta-metrics, we experimentally demonstrate AutoLibra’s ability to induce more concrete agent evaluation metrics than the ones proposed in previous agent evaluation benchmarks and discover new metrics to analyze agents. We also present two applications of AutoLibra in agent improvement: First, we show that AutoLibra serve human prompt engineers for diagonalize agent failures and improve prompts iterative. Moreover, we find that AutoLibra can induce metrics for automatic optimization for agents, which makes agents improve through self-regulation. Our results suggest that AutoLibra is a powerful task-agnostic tool for evaluating and improving language agents.

Evaluating the reasoning ability of language models (LMs) is complicated by their extensive parametric world knowledge, where benchmark performance often reflects factual recall rather than genuine reasoning. Existing datasets and approaches (e.g., temporal filtering, paraphrasing, adversarial substitution) cannot cleanly separate the two. We present SynthWorlds, a framework that disentangles task reasoning complexity from factual knowledge. In SynthWorlds, we construct parallel corpora representing two worlds with identical interconnected structure: a real-mapped world, where models may exploit parametric knowledge, and a synthetic-mapped world, where such knowledge is meaningless. On top of these corpora, we design two mirrored tasks as case studies: multi-hop question answering and page navigation, which maintain equal reasoning difficulty across worlds. Experiments in parametric-only (e.g., closed-book QA) and knowledge-augmented (e.g., retrieval-augmented) LM settings reveal a persistent *knowledge advantage gap*, defined as the performance boost models gain from memorized parametric world knowledge. Knowledge acquisition and integration mechanisms reduce but do not eliminate this gap, highlighting opportunities for system improvements. Fully automatic and scalable, SynthWorlds provides a controlled environment for evaluating LMs in ways that were previously challenging, enabling precise and testable comparisons of reasoning and memorization.

Agents powered by large language models (LLMs) are increasingly adopted in the software industry, contributing code as collaborators or even autonomous developers. As their presence grows, it becomes important to assess the current boundaries of their coding abilities. Existing agentic coding benchmarks, however, cover a limited task scope, e.g., bug fixing within a single pull request (PR), and often rely on non-executable evaluations or lack an automated approach for continually updating the evaluation coverage. To address such issues, we propose FeatureBench, a benchmark designed to evaluate agentic coding performance in end-to-end, feature-oriented software development. FeatureBench incorporates an execution-based evaluation protocol and a scalable test-driven method that automatically derives tasks from code repositories with minimal human effort. By tracing from unit tests along a dependency graph, our approach can identify feature-level coding tasks spanning multiple commits and PRs scattered across the development timeline, while ensuring the proper functioning of other features after the separation. Using this framework, we curated 200 challenging evaluation tasks and 3825 executable environments from 24 open-source repositories in the first version of our benchmark. Empirical evaluation reveals that the state-of-the-art agentic model, such as Claude 4.5 Opus, which achieves a 74.4% resolved rate on SWE-bench, succeeds on only 11.0% of tasks, opening new opportunities for advancing agentic coding. Moreover, benefiting from our automated task collection toolkit, FeatureBench can be easily scaled and updated over time to mitigate data leakage. The inherent verifiability of constructed environments also makes our method potentially valuable for agent training.

As Computer-Use Agents (CUAs) proliferate and grow increasingly capable, evaluation has become more challenging: static, manually curated benchmarks are narrow in domain, contamination-prone, and environment-heavy, and they diverge substantially from user-driven, real-world evaluation. We present Computer Agent Arena, an open-source platform for head-to-head CUA evaluation and a dynamic methodology that converts human preferences into structured feedback in realistic environments. The system (i) simulates real-world computer use via cloud-hosted, diverse, and dynamic environment initializations and customizations; (ii) ensures authentic, fair comparison by faithfully reproducing open-source CUAs and executing anonymously in matched, controlled environments; and (iii) extends evaluation beyond pairwise preference and correctness to capability- and behavior-oriented signals. Across 2,201 high-quality votes over 12 agents—spanning multi-app interactions, ambiguous instructions, and open-ended queries—we observe striking ranking reversals relative to static benchmarks. Further analysis shows that overall correctness mainly drives human preference; beyond that, agent-human interaction and self-correction boost user preference, even when overall task completion is comparable. Our error analysis reveals agent behavior errors, such as long-horizon memory and fine-grained action failures that static benchmarks fail to evaluate. We also contrast pure GUI agents with universal digital agents capable of tool use and coding, and discuss the trade-offs of these different design philosophies. We open source the full platform, collected dataset, and code of Computer Agent Arena to support future research on the evaluation and development of CUA.

In recent years, robotic manipulation policies have made substantial progress. However, evaluating these policies typically requires large-scale sampling in simulation benchmarks, leading to high time costs. Moreover, existing evaluation pipelines are usually fixed, do not account for user needs, and report only a single scalar score, lacking interpretability. In contrast, human experts can quickly form an intuitive impression of a policy’s capabilities from just a handful of executions. We therefore propose ManipEvalAgent, an efficient, promptable, and dynamically multi-round evaluation framework for robotic manipulation policies. The framework conducts small-batch, multi-round evaluations and adaptively plans subsequent evaluation steps based on intermediate observations from each round. Via code generation, it constructs tasks and evaluation functions within simulator. By generating evaluation functions and leveraging vision–language models (VLMs) for video understanding, ManipEvalAgent provides user-instruction-centric, fine-grained analysis. Our approach offers three key advantages: (1) efficiency, no need for massive sampling; (2) promptable, planning the evaluation process according to user queries; and (3) interpretability, providing diagnostic text that goes beyond a single score. Across multiple settings, our evaluation method significantly shortens the overall time compared with traditional simulation benchmarks, while reaching conclusions comparable to those from large-scale simulation benchmarks.

Does continued scaling of large language models (LLMs) yield diminishing returns? In this work, we show that short-task benchmarks may give an illusion of slowing progress, as even marginal gains in single-step accuracy can compound into exponential improvements in the length of tasks a model can successfully complete. Then, we argue that failures of LLMs when simple tasks are made longer arise from mistakes in execution, rather than an inability to reason. So, we propose isolating execution capability, by explicitly providing the knowledge and plan needed to solve a long-horizon task. First, we find that larger models can correctly execute significantly more turns even when small models have near-perfect single-turn accuracy. We then observe that the per-step accuracy of models degrades as the number of steps increases. This is not just due to long-context limitations---curiously, we observe a self-conditioning effect---models become more likely to make mistakes when the context contains their errors from prior turns. Self-conditioning does not reduce by just scaling the model size. But, we find that thinking mitigates self-conditioning, and also enables execution of much longer tasks in a single turn. We conclude by benchmarking frontier thinking models on the length of tasks they can execute in a single turn. Overall, by focusing on the ability to execute, we hope to reconcile debates on how LLMs can solve complex reasoning problems yet fail at simple tasks when made longer, and highlight the massive benefits of scaling model size and sequential test-time compute for long-horizon tasks.

Large language models split into two families: reasoning-centric LLMs, which strengthen internal chain-of-thought reasoning but cannot invoke external tools, and agentic LLMs, which learn to interact with environments and leverage tools but often lag in deep reasoning. This divide arises from fundamentally different training objectives, leading to mismatched strengths and inefficiency on simple queries, where both families tend to overthink or over-call tools. In this work, we present Adaptive Agent Foundation Model (A$^2$FM), a unified framework that follows a route-then-align principle: the model first learns task-aware routing and then aligns mode-specific trajectories under a shared backbone. To address the inefficiency gap, we introduce a third instant mode that handles simple queries directly, preventing unnecessary reasoning or tool calls while complementing the agentic and reasoning modes. To jointly enhance accuracy and efficiency, we propose Adaptive Policy Optimization (APO), which enforces adaptive sampling across modes and applies a cost-regularized reward. On the 32B scale, A$^2$FM achieves 13.4\% on BrowseComp, 70.4\% on AIME25, and 16.7\% on HLE, setting new SOTA among comparable models and performing competitively with frontier LLMs across agentic, reasoning, and general benchmarks. Notably, the adaptive execution achieves a cost of pass of only \$0.00487 per correct answer—cutting cost by 45.2\% relative to reasoning and 33.5\% relative to agentic, thus delivering substantially higher cost efficiency while maintaining comparable accuracy.

Although current large Vision-Language Models (VLMs) have advanced in multimodal understanding and reasoning, their fundamental perceptual and reasoning abilities remain limited. Specifically, even on simple jigsaw tasks, existing VLMs perform near randomly, revealing deficiencies in core perception and reasoning capabilities. While high-quality vision-language data can enhance these capabilities, its scarcity and limited scalability impose significant constraints. To address this, we propose AGILE, an Agentic jiGsaw Interaction Learning for Enhancing visual perception and reasoning in VLMs. AGILE formulates jigsaw solving as an interactive process, enabling the model to progressively engage with the environment. At each step, the model generates executable code to perform an action based on the current state, while the environment provides fine-grained visual feedback to guide task completion. Through this iterative cycle of observation and interaction, the model incrementally improves its perceptual and reasoning capabilities via exploration and feedback. Experimental results show that AGILE not only substantially boosts performance on jigsaw tasks of varying complexity (e.g., increasing accuracy from 9.5\% to 82.8\% under the $2 \times 2$ setting) but also demonstrates strong generalization across 9 general vision tasks, achieving an average improvement of 3.1\%. These results indicate notable enhancements in both perceptual and reasoning abilities. This work opens a new avenue for advancing reasoning and generalization in multimodal models and provides an efficient, scalable solution to the scarcity of multimodal reinforcement learning data. Code and datasets will be released soon.

Graphical user interface (GUI) agents autonomously complete tasks across platforms (\eg, Linux) by sequentially decomposing user instructions into action proposals that iteratively interact with visual elements in the evolving environment. However, two main challenges arise: i) planning (\ie, the action proposal sequence) under expansive action space, where selecting an appropriate plan is non-trivial, as many valid ones may exist; ii) accurately grounding actions in complex and high-resolution interfaces, \ie, precisely interacting with visual targets. This paper investigates the aforementioned challenges with our \textbf{G}UI \textbf{T}est-time Scaling \textbf{A}gent, namely \name. First, we conduct test-time scaling to select the most appropriate action proposal: at each step, multiple candidate proposals are sampled and evaluated and selected by a judge model. It trades off computation for better decision quality by concurrent sampling. Second, we propose a model that improves grounding of the selected action proposals to its corresponding visual elements. Our key insight is that reinforcement learning (RL) facilitates grounding through inherent objective alignments, rewarding successful clicks on interface elements. Experimentally, \name achieves state-of-the-art performance on both grounding and agent task execution benchmarks.

Despite the rapid growth of machine learning research, corresponding code implementations are often unavailable, making it slow and labor-intensive for researchers to reproduce results and build upon prior work. In the meantime, recent Large Language Models (LLMs) excel at understanding scientific documents and generating high-quality code. Inspired by this, we introduce PaperCoder, a multi-agent LLM framework that transforms machine learning papers into operational code repositories. PaperCoder operates in three stages: planning, where it constructs a high-level roadmap, designs the system architecture with diagrams, identifies file dependencies, and generates configuration files; analysis, which focuses on interpreting implementation-specific details; and generation, where modular, dependency-aware code is produced. Moreover, each phase is instantiated through a set of specialized agents designed to collaborate effectively across the pipeline. We then evaluate PaperCoder on generating code implementations from machine learning papers based on both model-based and human evaluations, particularly from the authors of those papers, with author-released repositories as ground truth if available. Our results demonstrate the effectiveness of PaperCoder in creating high-quality, faithful implementations. Furthermore, it consistently shows strengths in the recently released PaperBench benchmark, surpassing strong baselines by substantial margins.

Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A$^2$Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed $\mathrm{AnsF1}$ reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A$^2$Search achieves new state-of-the-art performance. With only a single rollout, A$^2$Search-7B yields an average $\mathrm{AnsF1}@1$ score of $48.4$% across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B ($46.2$%). Extensive analyses further show that A$^2$Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems.

Web search is a cornerstone for deep research agents, enabling them to acquire and reason over knowledge beyond static corpora. Yet most existing systems follow rigid ReAct-style tool chains locked into fixed workflow structures, which hinders their ability to flexibly handle diverse query types and tool-use strategies. We introduce $\textbf{FlowSearcher}$, a novel web search framework built on agentic workflow synthesis. FlowSearcher decomposes queries into sub-goals, each orchestrated by a tailored workflow graph that adapts the depth and order of tool use, giving the system structural flexibility to handle diverse sub-goals ranging from simple lookups and focused navigation to multi-hop information synthesis. Complementing this, a hierarchical memory distills past workflows into structured experience, providing reusable context that improves orchestration and guides tool use on new queries. This shift from reactive tool calls to memory-driven workflow design and execution marks a principled step toward deliberative web research. Empirical results on GAIA, BrowseComp, and GPQA show that our memory-driven, training-free workflow synthesis consistently matches or exceeds the performance of RLHF-trained systems, pointing toward a new direction of agent design grounded in memory-enhanced structural planning rather than parameter fine-tuning.

With an ever-growing zoo of LLMs and benchmarks, the need to orchestrate multiple models for improved task performance has never been more pressing. While frameworks like Mixture-of-Agents (MoA) attempt to coordinate LLMs, they often fall short in terms of (1) selecting relevant agents, (2) facilitating effective intra-agent communication, and (3) integrating responses efficiently. In this work, we propose Graph-of-Agents (GoA), a new graph-based framework for modeling multi-agent LLM communication. Our approach begins with node sampling, selecting only the most relevant agents by leveraging model cards that summarize each model’s domain, task specialization, and other characteristics. Next, we construct edges between the selected agents by evaluating their responses against one another to determine relevance ordering. Directed message passing is then performed from highly relevant agents to less relevant ones to enhance their responses, followed by reverse message passing to refine the original responses of the more relevant agents. Finally, the updated responses are aggregated via graph-based pooling (e.g., max or mean pooling) to produce a single, unified answer. We evaluate GoA on diverse multi-domain benchmarks (MMLU, MMLU-Pro, GPQA) and domain-specific benchmarks (MATH, HumanEval, MedMCQA), with an agent pool of 6 LLMs spanning multiple domains. Surprisingly, GoA achieves superior performance18 using only 3 selected agents, outperforming recent multi-agent LLM baselines that utilize all 6 agents simultaneously. By adopting a graph structure, GoA offers both scalability and effectiveness through structured message passing—positioning it as a strong candidate for navigating the challenges of the ever-growing LLM zoo.

Vision-Language Models (VLMs) are increasingly deployed as autonomous agents to navigate mobile Graphical User Interfaces (GUIs). However, their operation within dynamic on-device ecosystems, which include notifications, pop-ups, and inter-app interactions, exposes them to a unique and underexplored threat vector: environmental injection. Unlike traditional prompt-based attacks that manipulate textual instructions, environmental injection contaminates the agent's visual perception by inserting adversarial UI elements, such as deceptive overlays or spoofed notifications, directly into the GUI. This bypasses textual safeguards and can derail agent execution, leading to privacy leakage, financial loss, or irreversible device compromise. To systematically evaluate this threat, we introduce GhostEI-Bench, the first benchmark dedicated to assessing mobile agents under environmental injection attacks within dynamic, executable environments. Moving beyond static image-based assessments, our benchmark injects adversarial events into realistic application workflows inside fully operational Android emulators, assessing agent performance across a range of critical risk scenarios. We also introduce a novel evaluation protocol where a judge LLM performs fine-grained failure analysis by reviewing the agent's action trajectory alongside the corresponding sequence of screenshots. This protocol identifies the precise point of failure, whether in perception, recognition, or reasoning. Our comprehensive evaluation of state-of-the-art agents reveals their profound vulnerability to deceptive environmental cues. The results demonstrate that current models systematically fail to perceive and reason about manipulated UIs. GhostEI-Bench provides an essential framework for quantifying and mitigating this emerging threat, paving the way for the development of more robust and secure embodied agents.

Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While recent studies document instances of LLM deception, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce a new simulation framework, LH-Deception, for a systematic, empirical quantification of deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. LH-Deception is designed as a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed-source and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal emergent, long-horizon phenomena, such as ``chains of deception", which are invisible to static, single-turn evaluations. Our findings provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.

Deep research becomes increasingly important as people seek to solve complex problems that require gathering and synthesizing information from diverse sources. A key capability in this process is agentic search, where an LLM-agent iteratively retrieves relevant information across multiple sources while performing multi-step reasoning. However, developing effective agentic search systems is challenging due to the lack of high-quality training data that reflects the complexity of real-world research tasks. To address this gap, we introduce InfoSeek, a novel data synthesis framework that conceptualizes agentic search as a Hierarchical Constraint Satisfaction Problem (HCSP), where solving a task requires satisfying layered constraints across multiple levels of sub-problems. InfoSeek employs a Diffusion–Retrospection process: in the diffusion phase, the framework expands outward from a seed webpage, generating constraints that connect to neighboring pages and forming an exploration tree; in the retrospection phase, a subtree is sampled and backtracking constraints are introduced, which are then blurred and integrated into an HCSP instance. As a generic framework, InfoSeek can be easily extended to other domains beyond web, facilitating ad-hoc optimization of deep research. To our knowledge, InfoSeek is the first publicly released framework in this area, complete with open-source code and well-curated datasets. Extensive experiments on diverse information-seeking benchmarks show that training on InfoSeek-generated data substantially improves agentic search performance, delivering significantly larger gains than traditional datasets across diverse model backends and training strategies, thereby validating the effectiveness of our approach.

Medical Large Vision-Language Models (Med-LVLMs) have shown strong potential in multimodal diagnostic tasks. However, existing single-agent models struggle to generalize across diverse medical specialties, limiting their performance. Recent efforts introduce multi-agent collaboration frameworks inspired by clinical workflows, where general practitioners (GPs) and specialists interact in a fixed sequence. Despite improvements, these static pipelines lack flexibility and adaptability in reasoning. To address this, we propose MMedAgent-RL, a reinforcement learning (RL)-based multi-agent framework that enables dynamic, optimized collaboration among medical agents. Specifically, we train two GP agents based on Qwen2.5-VL via RL: the triage doctor learns to assign patients to appropriate specialties, while the attending physician integrates the judgments from multi-specialists and its own knowledge to make final decisions. To address the inconsistency in specialist outputs, we introduce a curriculum learning (CL)-guided RL strategy with dynamic entropy regulation, progressively teaching the attending physician to balance between imitating specialists and correcting their mistakes. Experiments on five medical VQA benchmarks demonstrate that MMedAgent-RL outperforms both open-source and proprietary Med-LVLMs. Notably, it achieves an average performance gain of 23.6\% over strong baselines.

Modern large language models (LLMs) such as GPT, Claude, and Gemini have transformed the way we learn, work, and communicate. Yet, their ability to produce highly human-like text raises serious concerns about misinformation and academic integrity, making it an urgent need for reliable algorithms to detect LLM-generated content. In this paper, we start by presenting a geometric approach to demystify rewrite-based detection algorithms, revealing their underlying rationale and demonstrating their generalization ability. Building on this insight, we introduce a novel rewrite-based detection algorithm that adaptively learns the distance between the original and rewritten text. Theoretically, we demonstrate that employing an adaptively learned distance function is more effective for detection than using a fixed distance. Empirically, we conduct extensive experiments with over 100 settings, and find that our approach demonstrates superior performance over baseline algorithms in the majority of scenarios. In particular, it achieves relative improvements from 57.8\% to 80.6\% over the strongest baseline across different target LLMs (e.g., GPT, Claude, and Gemini). A python implementation of our proposal is publicly available at https://github.com/Mamba413/L2D.

AI agents have significant potential to reshape cybersecurity, making a thorough assessment of their capabilities critical. However, existing evaluations fall short, because they are based on small-scale benchmarks and only measure static outcomes, failing to capture the full, dynamic range of real-world security challenges. To address these limitations, we introduce CyberGym, a large-scale benchmark featuring 1,507 real-world vulnerabilities across 188 software projects. Adjustable to different vulnerability analysis settings, CyberGym primarily tasks agents with generating a proof-of-concept test that reproduces a vulnerability, given only its text description and the corresponding codebase. Our extensive evaluation highlights that CyberGym effectively differentiates agents' and models' cybersecurity capabilities. Even the top-performing combinations only achieve a ~20% success rate, demonstrating the overall difficulty of CyberGym. Beyond static benchmarking, we show that CyberGym leads to the discovery of 35 zero-day vulnerabilities and 17 historically incomplete patches. These results underscore that CyberGym is not only a robust benchmark for measuring AI's progress in cybersecurity but also a platform for creating direct, real-world security impact.