NeurIPS 2025 Papers

Accurate temporal prediction is the bridge between comprehensive scene understanding and embodied artificial intelligence. However, predicting multiple fine-grained states of scene at multiple temporal scales is difficult for vision-language models. We formalize the Multi‐Scale Temporal Prediction (MSTP) task in general and surgical scene by decomposing multi‐scale into two orthogonal dimensions: the temporal scale, forecasting states of human and surgery at varying look‐ahead intervals, and the state scale, modeling a hierarchy of states in general and surgical scene. For instance in general scene, states of contacting relationship are finer-grained than states of spatial relationship. For instance in surgical scene, medium‐level steps are finer‐grained than high‐level phases yet remain constrained by their encompassing phase. To support this unified task, we introduce the first MSTP Benchmark, featuring synchronized annotations across multiple state scales and temporal scales. We further propose a novel method, Incremental Generation and Multi‐agent Collaboration (IG-MC), which integrates two key innovations. Firstly, we propose an plug-and-play incremental generation to keep high-quality temporal prediction that continuously synthesizes up-to-date visual previews at expanding temporal scales to inform multiple decision-making agents, ensuring decision content and generated visuals remain synchronized and preventing performance degradation as look‐ahead intervals lengthen. Secondly, we propose a decision‐driven multi‐agent collaboration framework for multiple states prediction, comprising generation, initiation, and multi‐state assessment agents that dynamically triggers and evaluates prediction cycles to balance global coherence and local fidelity. Extensive experiments on the MSTP Benchmark in general and surgical scene show that IG‐MC is a generalizable plug-and-play method for MSTP, demonstrating the effectiveness of incremental generation and the stability of decision‐driven multi‐agent collaboration.

Mathematical reasoning in large language models has been successfully incentivized through reinforcement learning with verifiable rewards, leading to improved one-shot precision. In this work, we turn our focus to the coding domain. Beyond one-shot precision, we highlight unit test generation as another key factor for enhancing coding ability, since accurate unit tests are essential for enabling self-checking and self-correction during inference. Traditional approaches for fine-tuning LLMs on unit test generation rely heavily on ground-truth code solutions in the training data. We propose CURE, a novel reinforcement learning framework with a dedicated reward design that co-evolves coding and unit test generation capabilities based on their interaction outcomes—without any ground-truth code as supervision. This approach enables flexible and scalable training and allows the unit tester to learn directly from the coder’s mistakes. Through extensive evaluations, we demonstrate that our CURE models, derived from base models of varying sizes, excel in both code generation and unit test generation. They naturally extend to downstream tasks such as test-time scaling—achieving a 6.2\% improvement over the base model—and agentic unit test generation, with a 25.1\% improvement. Our 4B model consistently outperforms Qwen3-4B while achieving 64.8\% inference efficiency in unit test generation. Notably, we also find that the CURE model can serve as an effective reward model for reinforcement learning on base models, even in the absence of any labeled supervision.

Rapidly improving large language models (LLMs) have the potential to assist in scientific progress. One critical skill in this endeavor is the ability to faithfully reproduce existing work. To evaluate the capability of AI agents to reproduce complex code in an active research area, we introduce the Automated LLM Speedrunning Benchmark, leveraging the research community's contributions to the $\textit{NanoGPT speedrun}$, a competition to train a GPT-2 model in the shortest time. Each of the 19 speedrun tasks provides the agent with the previous record's training script, optionally paired with one of three hint formats, ranging from pseudocode to paper-like descriptions of the new record's improvements. Records execute quickly by design and speedrun improvements encompass diverse code-level changes, ranging from high-level algorithmic advancements to hardware-aware optimizations. These features make the benchmark both accessible and realistic for the frontier problem of improving LLM training. We find that recent frontier reasoning LLMs combined with SoTA scaffolds struggle to reimplement already-known innovations in our benchmark, even when given detailed hints. Our benchmark thus provides a simple, non-saturated measure of an LLM's ability to automate scientific reproduction, a necessary (but not sufficient) skill for an autonomous research agent.

We study the selection of agents based on mutual nominations, a theoretical problem with many applications from committee selection to AI alignment. As agents both select and are selected, they may be incentivized to misrepresent their true opinion about the eligibility of others to influence their own chances of selection. Impartial mechanisms circumvent this issue by guaranteeing that the selection of an agent is independent of the nominations cast by that agent. Previous research has established strong bounds on the performance of impartial mechanisms, measured by their ability to approximate the number of nominations for the most highly nominated agents. We study to what extent the performance of impartial mechanisms can be improved if they are given a prediction of a set of agents receiving a maximum number of nominations. Specifically, we provide bounds on the consistency and robustness of such mechanisms, where consistency measures the performance of the mechanisms when the prediction is correct and robustness its performance when the prediction is incorrect. For the general setting where up to $k$ agents are to be selected and agents nominate any number of other agents, we give a mechanism with consistency $1-O\big(\frac{1}{k}\big)$ and robustness $1-\frac{1}{e}-O\big(\frac{1}{k}\big)$. For the special case of selecting a single agent based on a single nomination per agent, we prove that $1$-consistency can be achieved while guaranteeing $\frac{1}{2}$-robustness. A close comparison with previous results shows that (asymptotically) optimal consistency can be achieved with little to no sacrifice in terms of robustness.

Adaptive learning, which requires the in-depth understanding of students' learning processes and rational planning of learning resources, plays a crucial role in intelligent education. However, how to effectively model these two processes and seamlessly integrate them poses significant implementation challenges for adaptive learning. As core learning resources, exercises have the potential to diagnose students' knowledge states during the learning processes and provide personalized learning recommendations to strengthen students' knowledge, thereby serving as a bridge to boost student-oriented adaptive learning. Therefore, we introduce a novel task called Knowledge-aware Exercise Generative Recommendation (KEGR). It aims to dynamically infer students' knowledge states from their past exercise responses and customizably generate new exercises. To achieve KEGR, we propose an adaptive multi-agent cooperation framework, called ExeGen, inspired by the excellent reasoning and generative capabilities of LLM-based AI agents. Specifically, ExeGen coordinates four specialized agents for supervision, knowledge state perception, exercise generation, and quality refinement through an adaptive loop workflow pipeline. More importantly, we devise two enhancement mechanisms in ExeGen: 1) A human-simulated knowledge perception mechanism mimics students' cognitive processes and generates interpretable knowledge state descriptions via demonstration-based In-Context Learning (ICL). In this mechanism, a dual-matching strategy is further designed to retrieve highly relevant demonstrations for reliable ICL reasoning. 2) An exercise generation-adversarial mechanism collaboratively refines exercise generation leveraging a group of quality evaluation expert agents via iterative adversarial feedback. Finally, a comprehensive evaluation protocol is carefully designed to assess ExeGen. Extensive experiments on real-world educational datasets and a practical deployment in college education demonstrate the effectiveness and superiority of ExeGen. The code is available at https://github.com/dsz532/exeGen.

Non-stationary multi-armed bandits (nsMAB) enable agents to adapt to changing environments by incorporating mechanisms to detect and respond to shifts in reward distributions, making them well-suited for dynamic settings. However, existing approaches typically assume that reward feedback is available at every round—an assumption that overlooks many real-world scenarios where feedback is limited. In this paper, we take a significant step forward by introducing a new model of *constrained feedback in non-stationary multi-armed bandits* (ConFee-nsMAB), where the availability of reward feedback is restricted. We propose the first prior-free algorithm—that is, one that does not require prior knowledge of the degree of non-stationarity—that achieves near-optimal dynamic regret in this setting. Specifically, our algorithm attains a dynamic regret of $\tilde {\mathcal{O}}({K^{1/3} V_T^{1/3} T }/{ B^{1/3}})$, where $T$ is the number of rounds, $K$ is the number of arms, $B$ is the query budget, and $V_T$ is the variation budget capturing the degree of non-stationarity.

The rapid progress of navigation, manipulation, and vision models has made mobile manipulators capable in many specialized tasks. However, the open-world mobile manipulation (OWMM) task remains a challenge due to the need for generalization to open-ended instructions and environments, as well as the systematic complexity to integrate high-level decision making with low-level robot control based on both global scene understanding and current agent state. To address this complexity, we propose a novel multi-modal agent architecture that maintains multi-view scene frames and agent states for decision-making and controls the robot by function calling. A second challenge is the hallucination from domain shift. To enhance the agent performance, we further introduce an agentic data synthesis pipeline for the OWMM task to adapt the VLM model to our task domain with instruction fine-tuning. We highlight our fine-tuned OWMM-VLM as the first dedicated foundation model for mobile manipulators with global scene understanding, robot state tracking, and multi-modal action generation in a unified model. Through experiments, we demonstrate that our model achieves SOTA performance compared to other foundation models including GPT-4o and strong zero-shot generalization in real world. The project page is at https://hhyhrhy.github.io/owmm-agent-project.

Agents based on large language models (LLMs) for machine learning engineering (MLE) can automatically implement ML models via code generation. However, existing approaches to build such agents often rely heavily on inherent LLM knowledge and employ coarse exploration strategies that modify the entire code structure at once. This limits their ability to select effective task-specific models and perform deep exploration within specific components, such as experimenting extensively with feature engineering options. To overcome these, we propose MLE-STAR, a novel approach to build MLE agents. MLE-STAR first leverages external knowledge by using a search engine to retrieve effective models from the web, forming an initial solution, then iteratively refines it by exploring various strategies targeting specific ML components. This exploration is guided by ablation studies analyzing the impact of individual code blocks. Furthermore, we introduce a novel ensembling method using an effective strategy suggested by MLE-STAR. Our experimental results show that MLE-STAR achieves medals in 64% of the Kaggle competitions on the MLE-bench, significantly outperforming the best alternative.

LLM-based agents have been widely applied as personal assistants, capable of memorizing information from user messages and responding to personal queries. However, there still lacks an objective and automatic evaluation on their memory capability, largely due to the challenges in constructing reliable questions and answers (QAs) according to user messages. In this paper, we propose MemSim, a Bayesian simulator designed to automatically construct reliable QAs from generated user messages, simultaneously keeping their diversity and scalability. Specifically, we introduce the Bayesian Relation Network (BRNet) and a causal generation mechanism to mitigate the impact of LLM hallucinations on factual information, facilitating the automatic creation of an evaluation dataset. Based on MemSim, we generate a dataset in the daily-life scenario, named MemDaily, and conduct extensive experiments to assess the effectiveness of our approach. We also provide a benchmark for evaluating different memory mechanisms in LLM-based agents with the MemDaily dataset.

Fine-tuning pre-trained Large Language Models (LLMs) for downstream tasks using First-Order (FO) optimizers presents significant computational challenges. Parameter-Efficient Fine-Tuning~(PEFT) methods have been proposed to address these challenges by freezing most model parameters and training only a small subset. While PEFT is efficient, it may not outperform full fine-tuning when high task-specific performance is required. Zeroth-Order (ZO) methods offer an alternative for fine-tuning the entire pre-trained model by approximating gradients using only the forward pass, thus eliminating the computational burden of back-propagation, % in first-order methods, but they converge painfully slowly and are very sensitive to the choice of task prompts. We bridge these worlds with Bilevel‑ZOFO, a penalty‑based bilevel formulation that treats adapter parameters as a lower‑level learner coupled to an upper‑level ZO optimizer of the full backbone. This double-loop optimization strategy only requires the gradient of the PEFT model and the forward pass of the base model. We provide theoretical convergence guarantees for Bilevel ZOFO. Empirically, we demonstrate that Bilevel-ZOFO significantly outperforms existing ZO methods, achieves 2–4$\times$ faster training, and reduces sensitivity to prompts. Bilevel-ZOFO also outperforms FO PEFT methods while maintaining similar memory efficiency. Additionally, we show its strong potential for meta learning.

Speculative decoding is widely adopted to reduce latency in large language model (LLM) inference by leveraging smaller draft models capable of handling diverse user tasks. However, emerging AI applications, such as LLM-based agents, present unique workload characteristics: instead of diverse independent requests, agentic frameworks typically submit repetitive inference requests, such as multi-agent pipelines performing similar subtasks or self-refinement loops iteratively enhancing outputs. These workloads result in long and highly predictable sequences, which current speculative decoding methods do not effectively exploit. To address this gap, we introduce \emph{SuffixDecoding}, a novel method that utilizes efficient suffix trees to cache long token sequences from prompts and previous outputs. By adaptively speculating more tokens when acceptance likelihood is high and fewer when it is low, SuffixDecoding effectively exploits opportunities for longer speculations while conserving computation when those opportunities are limited. Evaluations on agentic benchmarks, including SWE-Bench and Text-to-SQL, demonstrate that SuffixDecoding achieves speedups of up to 3.9$\times$, outperforming state-of-the-art methods -- 2.2$\times$ faster than model-based approaches like EAGLE-2/3 and 1.6$\times$ faster than model-free approaches such as Token Recycling. SuffixDecoding is open-sourced.

Recent research on Reasoning of Large Language Models (LLMs) has sought to further enhance their performance by integrating meta-thinking—enabling models to monitor, evaluate, and control their reasoning processes for more adaptive and effective problem-solving. However, current single-agent work lacks a specialized design for acquiring meta-thinking, resulting in low efficacy. To address this challenge, we introduce Reinforced Meta-thinking Agents (ReMA), a novel framework that leverages Multi-Agent Reinforcement Learning (MARL) to elicit meta-thinking behaviors, encouraging LLMs to think about thinking. ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions. Through iterative reinforcement learning with aligned objectives, these agents explore and learn collaboration, leading to improved generalization and robustness. Empirical results from single-turn experiments demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks, including competitive-level mathematical benchmarks and LLM-as-a-Judge benchmarks. Additionally, we further extend ReMA to multi-turn interaction settings, leveraging turn-level ratio and parameter sharing to improve efficiency. Comprehensive ablation studies further illustrate the evolving dynamics of each distinct agent, providing valuable insights into how the meta-thinking reasoning process enhances the reasoning capabilities of LLMs.

Test-time scaling in agentic tasks often relies on generating long reasoning traces ("think" more) before acting, but this does not allow agents to acquire new information from the environment or adapt behavior over time. In this work, we propose scaling test-time interaction, an untapped dimension for test-time scaling that increases the agent's interaction horizon to enable rich behaviors such as exploration, backtracking, and dynamic re-planning within a single rollout. To demonstrate the promise of this scaling dimension, we situate our study in the domain of web agents. We first show that even prompting-based interaction scaling can improve task success on web benchmarks non-trivially. Building on this, we introduce TTI, a curriculum-based online reinforcement learning (RL) approach that trains agents by adaptively adjusting their interaction lengths during rollout. Using a Gemma 3 12B model, TTI sets a new state-of-the-art among open-source agents trained on public data on WebVoyager and WebArena. Case studies further reveal that TTI enables agents to balance exploration and exploitation adaptively. Our results establish interaction scaling as a powerful, complementary axis to scaling per-action compute, offering new avenues for training robust and adaptive agents.

Large language models (LLMs) have achieved significant performance progress in various natural language processing applications. However, LLMs still struggle to meet the strict requirements for accuracy and reliability in the medical field and face many challenges in clinical applications. Existing clinical diagnostic evaluation benchmarks for evaluating medical agents powered by LLMs have severe limitations. To address these limitations, we introduce ClinicalLab, a comprehensive clinical diagnosis agent alignment suite. ClinicalLab includes ClinicalBench, an end-to-end multi-departmental clinical diagnostic evaluation benchmark for evaluating medical agents and LLMs. ClinicalBench is based on real cases that cover 24 departments and 150 diseases. We ensure that ClinicalBench does not have data leakage. ClinicalLab also includes four novel metrics (ClinicalMetrics) for evaluating the effectiveness of LLMs in clinical diagnostic tasks. We evaluate 17 general and medical-domain LLMs and find that their performance varies significantly across different departments. Based on these findings, in ClinicalLab, we propose ClinicalAgent, an end-to-end clinical agent that aligns with real-world clinical diagnostic practices. We systematically investigate the performance and applicable scenarios of variants of ClinicalAgent on ClinicalBench. Our findings demonstrate the importance of aligning with modern medical practices in designing medical agents.

Foundation models are increasingly becoming better autonomous programmers, raising the prospect that they could also automate dangerous offensive cyber‑operations. Current frontier model audits probe the cybersecurity risks of such agents, but most fail to account for the degrees of freedom available to adversaries in the real world. In particular, with strong verifiers and financial incentives, agents for offensive cybersecurity are amenable to iterative improvement by would-be adversaries. We argue that assessments should take into account an expanded threat model in the context of cybersecurity, emphasizing the varying degrees of freedom that an adversary may possess in _stateful_ and _non-stateful_ environments within a fixed compute budget. We show that even with a relatively small compute budget (8 H100 GPU Hours in our study), adversaries can improve an agent's cybersecurity capability on InterCode CTF by more than 40\% relative to the baseline---without any external assistance. These results highlight the need to evaluate agents' cybersecurity risk in a dynamic manner, painting a more representative picture of risk.

Recently, test-time scaling of Large Language Models (LLMs) has emerged as a practical alternative to parameter and data scaling. Reasoning tasks often require large-scale, RLVR-based LLMs, while more economical LLMs can handle simpler tasks. Routing an LLM tailored to *suitability* (*i.e.*, capability and cost) ensures usability and efficiency. We introduce LLMRec, which routes the most suitable LLM to the user query without pre-inference on the candidate LLM zoo. It pioneeringly reframes the LLM routing problem as a comprehensive recommendation system (RecSys) task. Our core insight is that an LLM's suitability for a query is a complex, latent signal equal to user-item preference. LLMRec systematically engineers features for candidate LLMs (intrinsic attributes and capability distributions), queries (general semantics and meta-dimensional info), and context (inference type, cost budgets). It also incorporates behavioral features to learn high-order interactions. LLMRec is designed to generalize to out-of-domain datasets and adapt to new LLMs as the model zoo evolves. We define the metric with the Pareto frontier under user-specified cost budgets. Across six datasets, LLMRec achieves an average cost reduction of over 38% while maintaining accuracy and consistently outperforming baselines in converging toward the Pareto frontier.

Embodied agents powered by large language models (LLMs), such as Voyager, promise open-ended competence in worlds such as Minecraft. However, when powered by open-weight LLMs they still falter on elementary tasks after domain-specific fine-tuning. We propose MindForge, a generative-agent framework for cultural lifelong learning through explicit perspective taking. We introduce three key innovations: (1) a structured theory of mind representation linking percepts, beliefs, desires, and actions; (2) natural inter-agent communication; and (3) a multi-component memory system. Following the cultural learning framework, we test MindForge in both instructive and collaborative settings within Minecraft. In an instructive setting with GPT-4, MindForge agents powered by open-weight LLMs significantly outperform their Voyager counterparts in basic tasks yielding $3\times$ more tech-tree milestones and collecting $2.3\times$ more unique items than the Voyager baseline. Furthermore, in fully collaborative settings, we find that the performance of two underachieving agents improves with more communication rounds, echoing the Condorcet Jury Theorem. MindForge agents demonstrate sophisticated behaviors, including expert-novice knowledge transfer, collaborative problem solving, and adaptation to out-of-distribution tasks through accumulated cultural experiences.

The visual world around us constantly evolves, from real-time news and social media trends to global infrastructure changes visible through satellite imagery and augmented reality enhancements. However, Multimodal Large Language Models (MLLMs), which automate many tasks, struggle to stay current, limited by the cutoff dates in their fixed training datasets. To quantify this stagnation, we introduce LiveVQA, the first-of-its-kind dataset featuring 107,143 samples and 12 categories data specifically designed to support research in both seeking and updating with live visual knowledge. Drawing from recent news articles, video platforms, and academic publications in April 2024-May 2025, LiveVQA enables evaluation of how models handle latest visual information beyond their knowledge boundaries and how current methods help to update them. Our comprehensive benchmarking of 17 state-of-the-art MLLMs reveals significant performance gaps on content beyond knowledge cutoff, and tool-use or agentic visual seeking framework drastically gain an average of 327% improvement. Furthermore, we explore parameter-efficient fine-tuning methods to update MLLMs with new visual knowledge. We dive deeply to the critical balance between adapter capacity and model capability when updating MLLMs with new visual knowledge. All the experimental dataset and source code are publicly available at: https://livevqa.github.io.

Natural language has long enabled human cooperation, but its lossy, ambiguous, and indirect nature limits the potential of collective intelligence. While machines are not subject to these constraints, most LLM-based multi-agent systems still rely solely on natural language, exchanging tokens or their embeddings. To go beyond language, we introduce a new paradigm, *thought communication*, which enables agents to interact directly mind-to-mind, akin to telepathy. To uncover these latent thoughts in a principled way, we formalize the process as a general latent variable model, where agent states are generated by an unknown function of underlying thoughts. We prove that, in a nonparametric setting without auxiliary information, both shared and private latent thoughts between any pair of agents can be identified. Moreover, the global structure of thought sharing, including which agents share which thoughts and how these relationships are structured, can also be recovered with theoretical guarantees. Guided by the established theory, we develop a framework that extracts latent thoughts from all agents prior to communication and assigns each agent the relevant thoughts, along with their sharing patterns. This paradigm naturally extends beyond LLMs to all modalities, as most observational data arise from hidden generative processes. Experiments on both synthetic and real-world benchmarks validate the theory and demonstrate the collaborative advantages of thought communication. We hope this work illuminates the potential of leveraging the hidden world, as many challenges remain unsolvable through surface-level observation alone, regardless of compute or data scale.