NeurIPS 2025 Papers

Test-time scaling (TTS) enhances the performance of large language models (LLMs) by allocating additional compute resources during inference. However, existing research primarily investigates TTS in single-stage tasks; while many real-world problems are multi-stage complex tasks, composed of a sequence of heterogeneous subtasks with each subtask requires LLM of specific capability. Therefore, we study a novel problem: the test-time compute-optimal scaling in multi-stage complex tasks, aiming to select suitable models and allocate budgets per subtask to maximize overall performance. TTS in multi-stage tasks introduces two fundamental challenges: (i) The combinatorial search space of model and budget allocations, combined with the high cost of inference, makes brute-force search impractical. (ii) The optimal model and budget allocations across subtasks are interdependent, increasing the complexity of the compute-optimal search. To address this gap, we conduct extensive pilot experiments on four tasks across six datasets, deriving three empirical insights characterizing the behavior of LLMs in multi-stage complex tasks. Informed by these insights, we propose AgentTTS, an LLM-agent-based framework that autonomously searches for compute-optimal allocations through iterative feedback-driven interactions with the execution environment. Experimental results demonstrate that AgentTTS significantly outperforms traditional and other LLM-based baselines in search efficiency, and shows improved robustness to varying training set sizes and enhanced interpretability.

The rapid advancement of Large Language Models (LLMs) has stimulated interest in multi-agent collaboration for addressing complex medical tasks. However, the practical advantages of multi-agent collaboration approaches remain insufficiently understood. Existing evaluations often lack generalizability, failing to cover diverse tasks reflective of real-world clinical practice, and frequently omit rigorous comparisons against both single-LLM-based and established conventional methods. To address this critical gap, we introduce MedAgentBoard, a comprehensive benchmark for the systematic evaluation of multi-agent collaboration, single-LLM, and conventional approaches. MedAgentBoard encompasses four diverse medical task categories: (1) medical (visual) question answering, (2) lay summary generation, (3) structured Electronic Health Record (EHR) predictive modeling, and (4) clinical workflow automation, across text, medical images, and structured EHR data. Our extensive experiments reveal a nuanced landscape: while multi-agent collaboration demonstrates benefits in specific scenarios, such as enhancing task completeness in clinical workflow automation, it does not consistently outperform advanced single LLMs (e.g., in textual medical QA) or, critically, specialized conventional methods that generally maintain better performance in tasks like medical VQA and EHR-based prediction. MedAgentBoard offers a vital resource and actionable insights, emphasizing the necessity of a task-specific, evidence-based approach to selecting and developing AI solutions in medicine. It underscores that the inherent complexity and overhead of multi-agent collaboration must be carefully weighed against tangible performance gains. All code, datasets, detailed prompts, and experimental results are open-sourced at [this link](https://medagentboard.netlify.app/).

A multi-agent system (MAS) enhances its capacity to solve complex natural language processing (NLP) tasks through collaboration among multiple agents, where consensus-seeking serves as a fundamental mechanism. However, existing consensus-seeking approaches typically rely on voting mechanisms to judge consensus, overlooking contradictions in system-internal beliefs that destabilize the consensus. Moreover, these methods often involve agents updating their results through indiscriminate collaboration with every other agent. Such uniform interaction fails to identify the optimal collaborators for each agent, hindering the emergence of a stable consensus. To address these challenges, we provide a theoretical framework for selecting optimal collaborators that maximize consensus stability. Based on the theorems, we propose the Belief-Calibrated Consensus Seeking (BCCS) framework to facilitate stable consensus via selecting optimal collaborators and calibrating the consensus judgment by system-internal beliefs. Experimental results on the MATH and MMLU benchmark datasets demonstrate that the proposed BCCS framework outperforms the best existing results by 2.23\% and 3.95\% of accuracy on challenging tasks, respectively. Our code and data are available at https://github.com/dengwentao99/BCCS.

The application of AI in psychiatric diagnosis faces significant challenges, including the subjective nature of mental health assessments, symptom overlap across disorders, and privacy constraints limiting data availability. To address these issues, we present MoodAngels, the first specialized multi-agent framework for mood disorder diagnosis. Our approach combines granular-scale analysis of clinical assessments with a structured verification process, enabling more accurate interpretation of complex psychiatric data. Complementing this framework, we introduce MoodSyn, an open-source dataset of 1,173 synthetic psychiatric cases that preserves clinical validity while ensuring patient privacy. Experimental results demonstrate that MoodAngels outperforms conventional methods, with our baseline agent achieving 12.3\% higher accuracy than GPT-4o on real-world cases, and our full multi-agent system delivering further improvements. Together, these contributions provide both an advanced diagnostic tool and a critical research resource for computational psychiatry, bridging important gaps in AI-assisted mental health assessment.

Agentic search such as Deep Research systems-where agents autonomously browse the web, synthesize information, and return comprehensive citation-backed answers-represents a major shift in how users interact with web-scale information. While promising greater efficiency and cognitive offloading, the growing complexity and open-endedness of agentic search have outpaced existing evaluation benchmarks and methodologies, which largely assume short search horizons and static answers. In this paper, we introduce Mind2Web 2, a benchmark of 130 realistic, high-quality, and long-horizon tasks that require real-time web browsing and extensive information synthesis, constructed with over 1000 hours of human labor. To address the challenge of evaluating time-varying and complex answers, we propose a novel Agent-as-a-Judge framework. Our method constructs task-specific judge agents based on a tree-structured rubric design to automatically assess both answer correctness and source attribution. We conduct a comprehensive evaluation of ten frontier agentic search systems and human performance, along with a detailed error analysis to draw insights for future development. The best-performing system, OpenAI Deep Research, can already achieve 50-70% of human performance while spending half the time, highlighting its great potential. Altogether, Mind2Web 2 provides a rigorous foundation for developing and benchmarking the next generation of agentic search systems.

Scientific embodied agents play a crucial role in modern laboratories by automating complex experimental workflows. Compared to typical household environments, laboratory settings impose significantly higher demands on perception of physical-chemical transformations and long-horizon planning, making them an ideal testbed for advancing embodied intelligence. However, its development has been long hampered by the lack of suitable simulator and benchmarks. In this paper, we address this gap by introducing LabUtopia, a comprehensive simulation and benchmarking suite designed to facilitate the development of generalizable, reasoning-capable embodied agents in laboratory settings. Specifically, it integrates i) LabSim, a high-fidelity simulator supporting multi-physics and chemically meaningful interactions; ii) LabScene, a scalable procedural generator for diverse scientific scenes; and iii) LabBench, a hierarchical benchmark spanning five levels of complexity from atomic actions to long-horizon mobile manipulation. LabUtopia supports 30 distinct tasks and includes more than 200 scene and instrument assets, enabling large-scale training and principled evaluation in high-complexity environments. We demonstrate that LabUtopia offers a powerful platform for advancing the integration of perception, planning, and control in scientific-purpose agents and provides a rigorous testbed for exploring the practical capabilities and generalization limits of embodied intelligence in future research. Project web page: https://rui-li023.github.io/labutopia-site/

Current Large Language Models (LLMs) are confronted with overwhelming information volume when comprehending long-form documents. This challenge raises the imperative of a cohesive memory module, which can elevate vanilla LLMs into autonomous reading agents. Despite the emergence of some heuristic approaches, a systematic design principle remains absent. To fill this void, we draw inspiration from Jean Piaget's Constructivist Theory, illuminating three traits of the agentic memory---structured schemata, flexible assimilation, and dynamic accommodation. This blueprint forges a clear path toward a more robust and efficient memory system for LLM-based reading comprehension. To this end, we develop CAM, a prototype implementation of Constructivist Agentic Memory that simultaneously embodies the structurality, flexibility, and dynamicity. At its core, CAM is endowed with an incremental overlapping clustering algorithm for structured memory development, supporting both coherent hierarchical summarization and online batch integration. During inference, CAM adaptively explores the memory structure to activate query-relevant information for contextual response, akin to the human associative process. Compared to existing approaches, our design demonstrates dual advantages in both performance and efficiency across diverse long-text reading comprehension tasks, including question answering, query-based summarization, and claim verification.

Large language models are quickly becoming the foundation for intelligent agents that are capable of using tools. However, training such agents is challenging because it requires human creation and annotation of a diverse set of tasks, tools, and evaluation criteria. In this paper, we propose the Self-Challenging Agent framework for training an agent on high-quality tasks that are generated by itself. The agent first plays the role of challenger and generates a task after interacting with the given tools. The tasks take the form of a novel general class of problems termed Code-as-Task, which are defined by an instruction, a verification function and solution and failure cases which serve as tests, allowing to filter only for high-quality tasks. The agent then takes an executor role and trains on those tasks with reinforcement learning using the evaluation feedback as a reward. We show our method improves the performance of Llama-3.1-8B-Instruct on two existing multi-turn tool-use agent benchmarks, M$^3$ToolEval and TauBench, with a two-fold average success rate increase, despite using only self-generated training data.

Foundation models have recently shown strong potential as web agents, capable of interpreting high-level instructions and interacting with complex web interfaces. However, existing training paradigms for these agents often rely on predefined task datasets and curated demonstrations, limiting their scalability, adaptability, and capacity for self-improvement. In this work, we introduce *Self-guided hierArchical exploration for Generalist wEb agents* (SAGE), a new training framework designed to support autonomous skill acquisition through self-guided hierarchical exploration. Our method introduces a three-tier exploration strategy: a pre-exploration phase to build structural understanding of web environments, a top-level exploration strategy to generate a self-evolving curriculum of tasks from easy to hard, and a low-level exploration mechanism that combines planning-based rollouts with step-wise learning to improve policy efficiency. Together, these components form a scalable, supervision-free framework for web agent training. Experimental results on WebVoyager and WebArena demonstrate that our method significantly outperforms prior approaches, enabling foundation model agents to learn complex web tasks with greater generalization and robustness.

Existing AutoML systems have advanced the automation of machine learning (ML); however, they still require substantial manual configuration and expert input, particularly when handling multimodal data. We introduce MLZero, a novel multi-agent framework powered by Large Language Models (LLMs) that enables end-to-end ML automation across diverse data modalities with minimal human intervention. A cognitive perception module is first employed, transforming raw multimodal inputs into perceptual context that effectively guides the subsequent workflow. To address key limitations of LLMs, such as hallucinated code generation and outdated API knowledge, we enhance the iterative code generation process with semantic and episodic memory. MLZero demonstrates superior performance on MLE-Bench Lite, outperforming all competitors in both success rate and solution quality, securing six gold medals. Furthermore, when evaluated on our Multimodal AutoML Agent Benchmark, which includes 25 more challenging tasks spanning diverse data modalities, MLZero outperforms the competing methods by a large margin with a success rate of 0.92 (+263.6\%) and an average rank of 2.28. Our approach maintains its robust effectiveness even with a compact 8B LLM, outperforming full-size systems from existing solutions.

Financial markets pose fundamental challenges for asset return prediction due to their high dimensionality, non-stationarity, and persistent volatility. Despite advances in large language models and multi-agent systems, current quantitative research pipelines suffer from limited automation, weak interpretability, and fragmented coordination across key components such as factor mining and model innovation. In this paper, we propose R&D-Agent for Quantitative Finance, in short R&D-Agent(Q), the first data-centric multi-agent framework designed to automate the full-stack research and development of quantitative strategies via coordinated factor-model co-optimization. R&D-Agent(Q) decomposes the quant process into two iterative stages: a Research stage that dynamically sets goal-aligned prompts, formulates hypotheses based on domain priors, and maps them to concrete tasks, and a Development stage that employs a code-generation agent, Co-STEER, to implement task-specific code, which is then executed in real-market backtests. The two stages are connected through a feedback stage that thoroughly evaluates experimental outcomes and informs subsequent iterations, with a multi-armed bandit scheduler for adaptive direction selection. Empirically, R&D-Agent(Q) achieves up to 2× higher annualized returns than classical factor libraries using 70% fewer factors, and outperforms state-of-the-art deep time-series models on real markets. Its joint factor–model optimization delivers a strong balance between predictive accuracy and strategy robustness. Our code is available at: https://github.com/microsoft/RD-Agent.

Trajectory modeling, which includes research on trajectory data pattern mining and future prediction, has widespread applications in areas such as life services, urban transportation, and public administration. Numerous methods have been proposed to address specific problems within trajectory modeling. However, the heterogeneity of data and the diversity of trajectory tasks make effective and reliable trajectory modeling an important yet highly challenging endeavor, even for domain experts. In this paper, we propose TrajAgent, a agent framework powered by large language models (LLMs), designed to facilitate robust and efficient trajectory modeling through automation modeling. This framework leverages and optimizes diverse specialized models to address various trajectory modeling tasks across different datasets effectively. In TrajAgent, we first develop UniEnv, an execution environment with a unified data and model interface, to support the execution and training of various models. Building on UniEnv, we introduce an agentic workflow designed for automatic trajectory modeling across various trajectory tasks and data. Furthermore, we introduce collaborative learning schema between LLM-based agents and small speciallized models, to enhance the performance of the whole framework effectively. Extensive experiments on four tasks using four real-world datasets demonstrate the effectiveness of TrajAgent in automated trajectory modeling, achieving a performance improvement of 2.38%-69.91% over baseline methods. The codes and data can be accessed via https://github.com/tsinghua-fib-lab/TrajAgent.

Retrieval-augmented generation (RAG) is widely utilized to incorporate external knowledge into large language models, thereby enhancing factuality and reducing hallucinations in question-answering (QA) tasks. A standard RAG pipeline consists of several components, such as query rewriting, document retrieval, document filtering, and answer generation. However, these components are typically optimized separately through supervised fine-tuning, which can lead to misalignments between the objectives of individual components and the overarching aim of generating accurate answers. Although recent efforts have explored using reinforcement learning (RL) to optimize specific RAG components, these approaches often focus on simple pipelines with only two components or do not adequately address the complex interdependencies and collaborative interactions among the modules. To overcome these limitations, we propose treating the complex RAG pipeline with multiple components as a multi-agent cooperative task, in which each component can be regarded as an RL agent. Specifically, we present MMOA-RAG\footnote{The code of MMOA-RAG is on \url{https://github.com/chenyiqun/MMOA-RAG}.}, \textbf{M}ulti-\textbf{M}odule joint \textbf{O}ptimization \textbf{A}lgorithm for \textbf{RAG}, which employs multi-agent reinforcement learning to harmonize all agents' goals toward a unified reward, such as the F1 score of the final answer. Experiments conducted on various QA benchmarks demonstrate that MMOA-RAG effectively boost the overall performance of the pipeline and outperforms existing baselines. Furthermore, comprehensive ablation studies validate the contributions of individual components and demonstrate MMOA-RAG can be adapted to different RAG pipelines and benchmarks.

Combinatorial optimization problems involving multiple agents are notoriously challenging due to their NP-hard nature and the necessity for effective agent coordination. Despite advancements in learning-based methods, existing approaches often face critical limitations, including suboptimal agent coordination, poor generalization, and high computational latency. To address these issues, we propose PARCO (Parallel AutoRegressive Combinatorial Optimization), a general reinforcement learning framework designed to construct high-quality solutions for multi-agent combinatorial tasks efficiently. To this end, PARCO integrates three key novel components: (1) transformer-based communication layers to enable effective agent collaboration during parallel solution construction, (2) a multiple pointer mechanism for low-latency, parallel agent decision-making, and (3) priority-based conflict handlers to resolve decision conflicts via learned priorities. We evaluate PARCO in multi-agent vehicle routing and scheduling problems, where our approach outperforms state-of-the-art learning methods, demonstrating strong generalization ability and remarkable computational efficiency. We make our source code publicly available to foster future research: https://github.com/ai4co/parco.

AI agents today are mostly siloed — they either retrieve and reason over vast amount of digital information and knowledge obtained online; or interact with the physical world through embodied perception, planning and action — but rarely both. This separation limits their ability to solve tasks that require integrated physical and digital intelligence, such as cooking from online recipes, navigating with dynamic map data, or interpreting real-world landmarks using web knowledge. We introduce \textsc{Embodied Web Agents}, a novel paradigm for AI agents that fluidly bridge embodiment and web-scale reasoning. To operationalize this concept, we first develop the \textsc{Embodied Web Agents} task environments, a unified simulation platform that integrates realistic 3D indoor and outdoor environments with functional web interfaces. Building upon this platform, we construct and release the \textsc{Embodied Web Agents} Benchmark, which encompasses a diverse suite of tasks including cooking, navigation, shopping, tourism, and geolocation — all requiring coordinated reasoning across physical and digital realms for systematic assessment of cross-domain intelligence. Experimental results reveal significant performance gaps between state-of-the-art AI systems and human capabilities, establishing both challenges and opportunities at the intersection of embodied cognition and web-scale knowledge access.

Test-time scaling improves large language models' (LLMs) performance by allocating more compute budget during inference. To achieve this, existing methods often require intricate modifications to prompting and sampling strategies. In this work, we introduce representation consistency (RC), a test-time scaling method for aggregating answers drawn from multiple candidate responses of an LLM regardless of how they were generated, including variations in prompt phrasing and sampling strategy. RC enhances answer aggregation by not only considering the number of occurrences of each answer in the candidate response set, but also the consistency of the model's internal activations while generating the set of responses leading to each answer. These activations can be either dense (raw model activations) or sparse (encoded via pretrained sparse autoencoders). Our rationale is that if the model's representations of multiple responses converging on the same answer are highly variable, this answer is more likely to be the result of incoherent reasoning and should be down-weighted during aggregation. Importantly, our method only uses cached activations and lightweight similarity computations and requires no additional model queries. Through experiments with four open-source LLMs and four reasoning datasets, we validate the effectiveness of RC for improving task performance during inference, with consistent accuracy improvements (up to 4\%) over strong test-time scaling baselines. We also show that consistency in the sparse activation signals aligns well with the common notion of coherent reasoning.

Reinforcement learning is widely adopted in post-training large language models, especially for reasoning-style tasks such as maths questions. However, as we show, most existing methods will provably fail to learn from questions that are too hard, where the model always fails, or too easy, where the model always succeeds. Much human effort is therefore spent continually producing datasets of questions of a suitable difficulty for state-of-the-art models. Given this, we consider how to algorithmically identify questions that allow for maximally efficient training. We introduce a method, LILO (Learnability Improves LLMs Optimally), that prioritises training on questions with high variance of success, known as learnability, and we provide theory proving LILO maximises the expected improvement of the model. We run a wide range of experiments over multiple base models, algorithms and reasoning datasets to demonstrate that LILO consistently improves final test accuracy and can yield a 3x reduction in the number of training steps required to reach it. We explore how questions with high learnability can be efficiently identified, and discuss how learnability can be scaled to produce LLM agents that autonomously and open-endedly expand the frontier of human knowledge.

Humans continuously infer the states, goals, and behaviors of others by perceiving their surroundings in dynamic, real-world social interactions. However, most Theory of Mind (ToM) benchmarks only evaluate static, text-based scenarios, which have a significant gap compared to real interactions. We propose the SoMi-ToM benchmark, designed to evaluate multi-perspective ToM in embodied multi-agent complex social interactions. This benchmark is based on rich multimodal interaction data generated by the interaction environment SoMi, covering diverse crafting goals and social relationships. Our framework supports multi-level evaluation: (1) first-person evaluation provides multimodal (visual, dialogue, action, etc.) input from a first-person perspective during a task for real-time state inference, (2) third-person evaluation provides complete third-person perspective video and text records after a task for goal and behavior inference. This evaluation method allows for a more comprehensive examination of a model's ToM capabilities from both the subjective immediate experience and the objective global observation. We constructed a challenging dataset containing 35 third-person perspective videos, 363 first-person perspective images, and 1225 expert-annotated multiple-choice questions (three options). On this dataset, we systematically evaluated the performance of human subjects and several state-of-the-art large vision-language models (LVLMs). The results show that LVLMs perform significantly worse than humans on SoMi-ToM: the average accuracy gap between humans and models is 40.1% in first-person evaluation and 26.4% in third-person evaluation. This indicates that future LVLMs need to further improve their ToM capabilities in embodied, complex social interactions.

The fundamental challenge of long video understanding, e.g., question answering, lies in the extensive number of frames, making it infeasible to densely understand the local details while comprehensively digest the global contexts, especially within a limited context length. To address this problem, our insight is to process short video segments individually and combine these segment-level analyses into a final response. This intuition is noted in the well-established MapReduce principle in big data processing and is naturally compatible with inference scaling at the system level. Motivated by this, we propose MR. Video (pronounced as "mister video"), a long video understanding framework adopting the MapReduce principle. We define the standard operations of MapReduce in a long video understanding context: the Map steps conduct independent and sequence-parallel dense perception on short video segments, covering local details, while the Reduce steps comprehensively aggregate the segment-level results into an answer with global contexts. Thanks to the low cost and convenience of building video agents, we instantiate such Map and Reduce operations as an effective video agent capable of attending to local details and global contexts. Based on such abilities, we further introduce two critical yet previously under-explored long video understanding designs: (a) consistent character/object names in the captions, benefiting the reasoning of actions and stories across long horizons; (b) question intention analysis, which changes the key-frame retrieval in previous video agents to localizing the relevant information via jointly reasoning the whole video contexts and questions. Our MR. Video achieves a >7% accuracy improvement on the challenging LVBench over state-of-the-art video agents and vision-language models (VLMs) and demonstrates a clear advantage on multiple long video benchmarks, highlighting the potential of the MapReduce principle. The code is at https://github.com/ziqipang/MR-Video}{https://github.com/ziqipang/MR-Video.