ICLR 2026 论文列表

Large language models (LLMs) show the promise in supporting scientific research implementation, yet their ability to generate correct and executable code remains limited. Existing works largely adopt one-shot settings, ignoring the iterative and feedback-driven nature of realistic workflows of scientific research development. To address this gap, we present RECODE, a benchmark of 102 tasks from research papers and repositories that evaluates LLMs through multi-turn interactions with human feedback. It includes structured instructions, unit tests, and a five-level feedback hierarchy to reflect realistic researcher–agent collaboration. We further present ReCodeAgent, a framework that integrates feedback into iterative code generation. Experimentswith leading LLMs, including GPT-5, Claude-Sonnet-4, DeepSeek-V3.1, and Gemini 2.5, show substantial performance gains with richer feedback, while also highlighting ongoing challenges in the generation of complex research code. RECODE establishes a foundation for developing adaptive, feedback-driven LLM agents in scientific research implementation.

While scaling individual Large Language Models (LLMs) has delivered remarkable progress, the next frontier lies in scaling collaboration through multi-agent systems (MAS). However, purely autonomous MAS remain ``closed-world'' systems, constrained by the static knowledge horizon of pre-trained models. This limitation makes them brittle on tasks requiring knowledge beyond training data, often leading to collective failure under novel challenges. To address this, we propose the Learning to Intervene via Metacognitive Adaptation (LIMA) framework, a principled paradigm for human--agent collaboration. LIMA trains agents to learn a metacognitive policy that governs when to solve problems autonomously and when to defer to a human expert. To operationalize this policy, we introduce Dual-Loop Policy Optimization (DLPO), which disentangles immediate decision-making from long-term capability growth. The inner loop applies Group Relative Policy Optimization (GRPO) with a cost-aware reward to optimize deferral decisions, while the outer loop implements continual learning, transforming expert feedback into high-quality supervised signals that strengthen the agent's reasoning ability. Experiments on challenging mathematical and problem-solving benchmarks show that LIMA, equipped with Dual-Loop Optimization, consistently outperforms state-of-the-art MAS, establishing a principled foundation for collaborative and continually improving agentic systems.

We introduce DRBench, a benchmark for evaluating AI agents on complex, open-ended deep research tasks in enterprise settings. Unlike prior benchmarks that focus on simple questions or web-only queries, DRBench evaluates agents on multi-step queries (for example, "What changes should we make to our product roadmap to ensure compliance with this standard?") that require identifying supporting facts from both the public web and private company knowledge base. Each task is grounded in realistic user personas and enterprise context, spanning a heterogeneous search space that includes productivity software, cloud file systems, emails, chat conversations, and the open web. Tasks are generated through a carefully designed synthesis pipeline with human-in-the-loop verification, and agents are evaluated on their ability to recall relevant insights, maintain factual accuracy, and produce coherent, well-structured reports. We release 100 deep research tasks across 10 domains, such as Sales, Cybersecurity, and Compliance. We demonstrate the effectiveness of DRBench by evaluating diverse DR agents across open- and closed-source models (such as GPT, Llama, and Qwen) and DR strategies, highlighting their strengths, weaknesses, and the critical path for advancing enterprise deep research.

\begin{abstract} Developing generalist large language models (LLMs) capable of complex biological reasoning is a central challenge in computational biology. While existing LLMs excel at predictive tasks like cell type annotation and logically-constrained problems, enabling open-ended and mechanistic reasoning remains a challenge. A promising direction is Reinforcement Learning from Verifiable Rewards (RLVR), which has been shown to significantly enhance complex reasoning in general domains like mathematics and code synthesis. However, its application in biology is hindered, as most biological outcomes are non-verifiable. For example, verifying a generated gene sequence is usually infeasible. In this paper, we introduce CellDuality, a self-supervised framework that enables LLM agents for robust reasoning in single-cell biology. Our framework is built on the principle of complementary task duality, a self-verification process that leverages a bidirectional reasoning loop. First, the model performs a forward reasoning task by predicting a biological outcome (e.g., a cell's response to a drug). Then, in a complementary inverse task, it must reason backward from its own prediction to reconstruct the initial conditions (e.g., the original drug perturbation). The fidelity of this reconstruction serves as an intrinsic reward signal, creating a feedback loop that enforces logical and biological consistency. We use these intrinsic rewards to align the base LLM via reinforcement learning, without requiring ground-truth verification labels. We demonstrate that CellDuality achieves state-of-the-art performance and provides coherent biological explanations across a diverse suite of single-cell reasoning tasks. Critically, on the challenging out-of-distribution perturbation prediction benchmark, our self-supervised approach significantly outperforms the standard fine-tuning baseline and narrows the performance gap to a supervised RLVR baseline. Our work showcases a new path toward scalable training of biological foundation models.

Large language models, employed as multiple agents that interact and collaborate with each other, have excelled at solving complex tasks. The agents are programmed with prompts that declare their functionality, along with the topologies that orchestrate interactions across agents. Designing prompts and topologies for multi-agent systems (MAS) is inherently complex. To automate the entire design process, we first conduct an in-depth analysis of the design space aiming to understand the factors behind building effective MAS. We reveal that prompts together with topologies play critical roles in enabling more effective MAS design. Based on the insights, we propose Multi-Agent System Search (MASS), a MAS optimization framework that efficiently exploits the complex MAS design space by interleaving its optimization stages, from local to global, from prompts to topologies, over three stages: 1) block-level (local) prompt optimization; 2) workflow topology optimization; 3) workflow-level (global) prompt optimization, where each stage is conditioned on the iteratively optimized prompts/topologies from former stages. We show that MASS-optimized multi-agent systems outperform a spectrum of existing alternatives by a substantial margin. Based on the MASS-found systems, we finally propose design principles behind building effective multi-agent systems.

To significantly advance the capabilities of open-source web agents, we present WebSailor-V2, a complete post-training pipeline encompassing data construction, Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL). Our methodology features two key innovations: (1) On the data front, we developed SailorFog-QA-2, a novel dataset built from a densely interconnected knowledge graph that introduces a wide variety of uncertainties beyond simple obfuscation, fostering more sophisticated reasoning. (2) For training, we engineered a dual-environment RL framework, combining a high-fidelity simulator for rapid, low-cost algorithmic iteration with a robust, managed real-world environment for stable final policy training, all integrated within a symbiotic data-policy feedback loop. Trained on the Qwen3-30B-A3B model, WebSailor-V2 achieves state-of-the-art results, scoring 35.3 on BrowseComp-EN, 44.1 on BrowseComp-ZH, and 30.6 on Humanity's Last Exam (HLE). Notably, our 30B-A3B MOE agent significantly outperforms all existing open-source agents and surpasses even the 671B DeepSeek-V3.1, demonstrating performance competitive with leading proprietary systems.

Structured images (e.g., charts and geometric diagrams) remain challenging for multimodal large language models (MLLMs), as perceptual slips can cascade into erroneous conclusions. Intermediate visual cues can steer reasoning; however, existing cue-based methods are constrained with low-fidelity image processing and linear, rigid reasoning patterns, limiting their effectiveness on complex structured-image tasks. In this paper, we propose PixelCraft, a novel multi-agent system for high-fidelity image processing and flexible visual reasoning on structured images. The system comprises a dispatcher, a planner, a reasoner, critics, and a set of visual tool agents. To achieve high-fidelity processing, we construct a high-quality corpus and fine-tune an MLLM into a grounding model, whose pixel-level localizations are integrated with traditional computer vision (CV) algorithms in tool agents. Building on this foundation, PixelCraft facilitates flexible visual reasoning through a dynamic three-stage workflow of tool selection, agent discussion, and self-criticism. Moreover, unlike prior linear reasoning patterns that simply append historical images, PixelCraft maintains an image memory to allow the planner to adaptively revisit earlier visual steps, explore alternative reasoning branches, and dynamically adjust the reasoning trajectory during discussion. Extensive experiments on challenging chart and geometry benchmarks demonstrate that PixelCraft significantly improves visual reasoning performance for advanced MLLMs, setting a new standard for structured image reasoning.

Current paradigms for training GUI agents are fundamentally limited by a reliance on either unsafe, non-reproducible live web interactions or costly, scarce human-crafted data and environments. We argue this focus on data volume overlooks a more critical factor: the efficiency of compressing a large language model's (LLM) latent knowledge into actionable agent behavior. We introduce WebFactory, a novel, fully automated closed-loop reinforcement learning pipeline for GUI agents, systematically compressing LLM-encoded internet intelligence into efficient, grounded actions. Our pipeline features a process of scalable environment synthesis → knowledge-aware task generation → LLM-powered trajectory collection → decomposed reward RL training → systematic agent evaluation. Remarkably, our agent demonstrates exceptional data efficiency and generalization. Trained on synthetic data from only 10 websites within WebFactory, it achieves performance comparable to GUI agents trained on same amount of human-annotated data from a much larger set of environments. This superior performance is consistent across our internal offline and online transferring benchmarks, where our agent also significantly outperforms the base foundation model. We further provide critical insights into the "embodiment potential" of different LLM foundations, offering a new axis for model evaluation. This work presents a scalable and cost-effective paradigm for transforming passive internet knowledge into active, grounded intelligence, marking a critical step towards general-purpose interactive agents.

While Large Language Models (LLMs) have demonstrated significant advancements in reasoning and agent-based problem-solving, current evaluation methodologies fail to adequately assess their capabilities: existing benchmarks either rely on closed-ended questions prone to saturation and memorization, or subjective comparisons that lack consistency and rigor. In this work, we introduce HeuriGym, an agentic framework designed for evaluating heuristic algorithms generated by LLMs for combinatorial optimization problems, characterized by clearly defined objectives and expansive solution spaces. HeuriGym empowers LLMs to propose heuristics, receive evaluative feedback via code execution, and iteratively refine their solutions. We evaluate nine state-of-the-art models on various problems across domains such as computer systems, logistics, and biology, exposing persistent limitations in tool use, planning, and adaptive reasoning. To quantify performance, we propose the Quality-Yield Index (QYI), a metric that captures both solution pass rate and quality. Even top models like GPT-o4-mini-high and Gemini-2.5-Pro attain QYI scores of only 0.6, well below the expert baseline of 1. Our open-source benchmark aims to guide the development of LLMs toward more effective and realistic problem-solving in scientific and engineering domains.

As LLMs become more agentic, a critical risk emerges: attackers can \emph{decompose} harmful goals into stateful, benign subtasks that trick LLM agents into executing them without realizing the harmful intent in the same context. The challenge lies in the existing shallow safety alignment techniques: they only detect harm in the immediate prompt and do not reason about long-range intent. We therefore propose adding an external monitor that observes the conversation at a higher level. To facilitate our study on monitoring decomposition attacks, we curate the largest and most diverse dataset, DecomposedHarm, with 4,634 tasks that can be assigned to LLM agents, including general agent tasks, text-to-image, and question-answering tasks, where each task has a benignly decomposed version. We verify our datasets by testing them on frontier models and show an 87\% attack success rate on average on GPT-4o. To defend in real‐time, we propose a lightweight sequential monitoring framework that cumulatively evaluates each sub‑prompt. We show that a carefully prompt-engineered lightweight monitor hits a 93\% defense success rate—outperforming strong baselines such as Llama-Guard-4 and o3-mini, while cutting costs by 90\% and latency by 50\%. Additionally, we show that even under adversarial pressure, combining decomposition attacks with massive random task injection and automated red teaming, our lightweight sequential monitors remain robust. Our findings suggest that guarding against stateful decomposition attacks is "surprisingly easy" with lightweight sequential monitors, enabling safety in real-world LLM agent deployment where expensive solutions are impractical.

While integrating tools like Code Interpreter and Search has significantly enhanced Large Language Model (LLM) reasoning in models like ChatGPT Agent and Gemini-Pro, practical guidance on optimal tool use is lacking. The core challenge is effectively combining textual reasoning, coding, and search for diverse questions. In this paper, we propose Tool-Use Mixture (TUMIX), an ensemble framework that runs multiple agents in parallel, each employing distinct tool-use strategies and answer paths. Agents in TUMIX iteratively share and refine responses based on the question and previous answers. In experiments, TUMIX achieves significant gains over state-of-the-art tool-augmented and test-time scaling methods, delivering an average accuracy improvement of up to 3.55\% over the best baseline on Gemini-2.5-Pro and Gemini-2.5-Flash across key reasoning benchmarks, with near-equal inference costs. We find that agent diversity and quality are crucial and can be enhanced by using LLMs to auto-optimize agent designs. Furthermore, TUMIX can halt refinement upon reaching sufficient confidence, preserving performance at only 49\% of the inference cost. Further scaling can achieve higher performance, albeit at a greater cost.

Execution-based feedback like unit testing is widely used in the development of coding agents through test-time scaling (TTS) and reinforcement learning (RL). This paradigm requires scalable and reliable collection of unit test cases to provide accurate feedback, and the resulting feedback is often sparse and cannot effectively distinguish between trajectories that are both successful or both unsuccessful. In contrast, execution-free feedback from reward models can provide more fine-grained signals without depending on unit test cases. Despite this potential, execution-free feedback for realistic software engineering (SWE) agents remains underexplored. Aiming to develop versatile reward models that are effective across TTS and RL, however, we observe that two verifiers with nearly identical TTS performance can nevertheless yield very different results in RL. Intuitively, TTS primarily reflects the model’s ability to select the best trajectory, but this ability does not necessarily generalize to RL. To address this limitation, we identify two additional aspects that are crucial for RL training: classification accuracy and calibration. We then conduct comprehensive controlled experiments to investigate how to train a robust reward model that performs well across these metrics. In particular, we analyze the impact of various factors such as training data scale, policy mixtures, and data source composition. Guided by these investigations, we introduce SWE-RM, an accurate and robust reward model adopting a mixture-of-experts architecture with 30B total parameters and 3B activated during inference. SWE-RM substantially improves SWE agents on both TTS and RL performance. For example, it increases the accuracy of Qwen3-Coder-Flash from 51.6% to 62.0%, and Qwen3-Coder-Max from 67.0% to 74.6% on SWE-Bench Verified using TTS, achieving new state-of-the-art performance among open-source models. On RL training, SWE-RM lifts the resolve rate of execution-based counterparts by 3 absolute points on SWE-Bench Verified.

Large Language Model (LLM) agents are reshaping the game industry, by enabling more intelligent and human-preferable characters. Yet, current game benchmarks fall short of practical needs: they lack evaluations of diverse LLM capabilities across various game genres, studies of agentic modules crucial for complex gameplay, and fine-tuning datasets to adapt pre-trained LLMs into gaming agents. To fill these gaps, we present Orak, a benchmark for training and evaluating LLM agents across 12 popular video games spanning all major genres. Using a plug-and-play interface built on Model Context Protocol (MCP), Orak supports systematic and reproducible studies of agentic modules in varied game scenarios. We further release a fine-tuning dataset of expert LLM gameplay trajectories spanning multiple genres, turning general LLMs into effective game agents. Orak offers a comprehensive evaluation framework, including game leaderboards, LLM battle arenas, and in-depth analyses of input modality, agentic strategies, and fine-tuning effects, establishing a foundation towards versatile gaming agents. Code is available at https://anonymous.4open.science/r/Orak-5013/.

Unknown unknowns (U2s) are deployment-time scenarios absent from development/testing. Unlike conventional anomalies, U2s are not out-of-distribution (OOD); they stem from changes in underlying system dynamics without a distribution shift from normal data. Thus, existing multi-variate time series anomaly detection (MTAD) methods—which rely on distribution-shift cues—are ill-suited for U2 detection. Specifically: (i) we show most anomaly datasets exhibit distribution shift between normal and anomalous data and therefore are not representative of U2s; (ii) we introduce eight U2 benchmarks where training data contain OOD anomalies but no U2s, while test sets contain both OOD anomalies and U2s; (iii) we demonstrate that state-of-the-art (SOTA) MTAD results often depend on impractical enhancements: point adjustment (PA) (uses ground truth to flip false negatives to true positives, inflating precision) and threshold learning with data leakage (TL) (tuning thresholds on test data and labels); (iv) with PA+TL, even untrained deterministic methods can match or surpass MTAD baselines; (v) without PA/TL, existing MTAD methods degrade sharply on U2 benchmarks. Finally, we present sparse model identification–enhanced anomaly detection (SPIE-AD), a model-recovery-and-conformance, zero-shot MTAD approach that outperforms baselines on all eight U2 benchmarks and on six additional real-world MTAD datasets—without PA or TL.

Large language models (LLMs) are emerging as powerful tools for scientific law discovery, a foundational challenge in AI-driven science. However, existing benchmarks for this task suffer from a fundamental methodological trilemma, forcing a trade-off between scientific relevance, scalability, and resistance to memorization. Furthermore, they oversimplify discovery as static function fitting, failing to capture the authentic scientific process of uncovering embedded laws through the interactive exploration of complex model systems. To address these critical gaps, we introduce **NewtonBench**, a benchmark comprising 324 scientific law discovery tasks across 12 physics domains. Our design mitigates the evaluation trilemma by using counterfactual law shifts - systematic alterations of canonical laws - to generate a vast suite of problems that are scalable, scientifically relevant, and memorization-resistant. Moreover, we elevate the evaluation from static function fitting to interactive model discovery, requiring agents to experimentally probe simulated complex systems to uncover hidden principles. Our extensive evaluation of 11 state-of-the-art LLMs reveals a clear but fragile capability for discovery in frontier models: this ability degrades precipitously with increasing system complexity and exhibits extreme sensitivity to observational noise. Notably, we uncover a paradoxical effect of tool assistance: providing a code interpreter can hinder more capable models by inducing a premature shift from exploration to exploitation, causing them to satisfice on suboptimal solutions. These results demonstrate that robust, generalizable discovery in complex, interactive environments remains the core challenge for the future of automated science. By providing a scalable, robust, and scientifically authentic testbed, NewtonBench offers a crucial tool for measuring true progress and guiding the development of next-generation AI agents capable of genuine scientific discovery.

Prevalent retrieval-based tool-use pipelines struggle with a dual semantic challenge: their retrievers often employ encoders that fail to capture complex semantics, while the Large Language Model (LLM) itself lacks intrinsic tool knowledge from its natural language pretraining. Generative methods offer a powerful alternative by unifying selection and execution, tasking the LLM to directly learn and generate tool identifiers. However, the common practice of mapping each tool to a unique new token introduces substantial limitations: it creates a scalability and generalization crisis, as the vocabulary size explodes and each tool is assigned a semantically isolated token. This approach also creates a semantic bottleneck that hinders the learning of collaborative tool relationships, as the model must infer them from sparse co-occurrences of monolithic tool IDs within a vast library. To address these limitations, we propose ToolWeaver, a novel generative tool learning framework that encodes tools into hierarchical sequences. This approach makes vocabulary expansion logarithmic to the number of tools. Crucially, it enables the model to learn collaborative patterns from the dense co-occurrence of shared codes, rather than the sparse co-occurrence of monolithic tool IDs. We generate these structured codes through a novel tokenization process designed to weave together a tool's intrinsic semantics with its extrinsic co-usage patterns. These structured codes are then integrated into the LLM through a generative alignment stage, where the model is fine-tuned to produce the hierarchical code sequences. Evaluation results with nearly 47,000 tools show that ToolWeaver significantly outperforms state-of-the-art methods, establishing a more scalable, generalizable, and semantically-aware foundation for advanced tool-augmented agents.

Large Language Model (LLM) pretraining, finetuning, and evaluation rely on input-space reconstruction and generative capabilities. Yet, it has been observed in vision that embedding-space training objectives, e.g., with Joint Embedding Predictive Architectures (JEPAs), are far superior to their input-space counterpart. That mismatch in how training is achieved between language and vision opens up a natural question: {\em can language training methods learn a few tricks from the vision ones?} The lack of JEPA-style LLM is a testimony of the challenge in designing such objectives for language. In this work, we propose a first step in that direction where we develop LLM-JEPA, a JEPA based solution for LLMs applicable both to finetuning and pretraining. Thus far, LLM-JEPA is able to outperform the standard LLM training objectives by a significant margin across models, all while being robust to overfiting. Those findings are observed across numerous datasets (NL-RX, GSM8K, Spider, RottenTomatoes) and various models from the Llama3, OpenELM, Gemma2 and Olmo families. Code: \url{https://github.com/galilai-group/llm-jepa}.

LLMs and Agents have achieved impressive progress in code generation, mathematical reasoning, and scientific discovery. However, existing benchmarks primarily measure correctness, overlooking the diversity of methods behind solutions. True innovation depends not only on producing correct answers but also on the originality of the approach. We present \textbf{InnoGym}, the first benchmark and framework designed to systematically evaluate the innovation potential of AI agents. InnoGym introduces two complementary metrics: performance gain, which measures improvement over the best-known solutions, and novelty, which captures methodological differences from prior approaches. The benchmark includes 18 carefully curated tasks from real-world engineering and scientific domains, each standardized through resource filtering, evaluator validation, and solution collection. In addition, we provide \textbf{iGym}, a unified execution environment for reproducible and long-horizon evaluations. Extensive experiments show that while some agents produce novel approaches, their lack of robustness limits performance gains. These results highlight a key gap between creativity and effectiveness, underscoring the need for benchmarks that evaluate both.

Embodied agents operating in multi-agent, partially observable, and decentralized environments must plan and act despite pervasive uncertainty about hidden objects and collaborators' intentions. Recent advances in applying Large Language Models (LLMs) to embodied agents have addressed many long-standing challenges, such as high-level goal decomposition and online adaptation. Yet, uncertainty is still primarily mitigated through frequent inter-agent communication. This incurs substantial token and time costs, and can disrupt established workflows, when human partners are involved. We introduce PCE, a Planner-Composer-Evaluator framework that converts the fragmented assumptions latent in LLM reasoning traces into a structured decision tree. Internal nodes encode environment assumptions and leaves map to actions; each path is then scored by scenario likelihood, goal-directed gain, and execution cost to guide rational action selection without heavy communication. Across two challenging multi-agent benchmarks (C-WAH and TDW-MAT) and three diverse LLM backbones, PCE consistently outperforms communication-centric baselines in success rate and task efficiency while showing comparable token usage. Ablation results indicate that the performance gains obtained by scaling model capacity or reasoning depth persist even when PCE is applied, while PCE consistently raises the baseline across both capacity and reasoning-depth scales, confirming that structured uncertainty handling complements both forms of scaling. A user study further demonstrates that PCE produces communication patterns that human partners perceive as more efficient and trustworthy. Together, these results establish a principled route for turning latent LLM assumptions into reliable strategies for uncertainty-aware planning.