NeurIPS 2025 Papers

The ability to selectively remove knowledge from LLMs is highly desirable. However, existing methods often struggle with balancing unlearning efficacy and retain model utility, and lack controllability at inference time to emulate base model behavior as if it had never seen the unlearned data. In this paper, we propose LUNAR, a novel unlearning method grounded in the Linear Representation Hypothesis and operates by redirecting the representations of unlearned data to activation regions that expresses its inability to answer. We show that contrastive features are not a prerequisite for effective activation redirection, and LUNAR achieves state-of-the-art unlearning performance and superior controllability. Specifically, LUNAR achieves between 2.9x and 11.7x improvement in the combined unlearning efficacy and model utility score (Deviation Score) across various base models and generates coherent, contextually appropriate responses post-unlearning. Moreover, LUNAR effectively reduces parameter updates to a single down-projection matrix, a novel design that significantly enhances efficiency by 20x and robustness. Finally, we demonstrate that LUNAR is robust to white-box adversarial attacks and versatile in real-world scenarios, including handling sequential unlearning requests.

Compiler auto-tuning optimizes pass sequences to improve performance metrics such as Intermediate Representation (IR) instruction count. Although recent advances leveraging Large Language Models (LLMs) have shown promise in automating compiler tuning, two significant challenges still remain: the absence of high-quality reasoning datasets for agents training, and limited effective interactions with the compilation environment. In this work, we introduce Compiler-R1, the first reinforcement learning (RL)-driven framework specifically augmenting LLM capabilities for compiler auto-tuning. Compiler-R1 features a curated, high-quality reasoning dataset and a novel two-stage end-to-end RL training pipeline, enabling efficient environment exploration and learning through an outcome-based reward. Extensive experiments across seven datasets demonstrate Compiler-R1 achieving an average 8.46\% IR instruction count reduction compared to opt -Oz, showcasing the strong potential of RL-trained LLMs for compiler optimization. Our code and datasets are publicly available at https://github.com/Panhaolin2001/Compiler-R1.

Recently, Sharma et al. (2024) suggested a method called LAyer- SElective-Rank reduction (LASER) which demonstrated that pruning high‑order components of carefully chosen LLM’s weight matrices can boost downstream accuracy—without any gradient‑based fine‑tuning. Yet LASER’s exhaustive, per‑matrix search (each requiring full‑dataset forward passes) makes it impractical for rapid deployment. We demonstrate that this overhead can be removed and find that: (i) Only a small, carefully chosen subset of matrices needs to be inspected—eliminating the layer‑by‑layer sweep, (ii) The gradient of each matrix’s singular values pinpoints which matrices merit reduction, (iii) Increasing the factorization search space by allowing matrices rows to cluster around multiple subspaces and then decomposing each cluster separately further reduces overfitting on the original training data and further lifts accuracy by up to 24.6 percentage points, and finally, (iv) we discover that evaluating on just 100 samples rather than the full training data—both for computing the indicative gradients and for measuring the final accuracy—suffices to further reduce the search time; we explain that as adaptation to downstream tasks is dominated by prompting style, not dataset size. As a results, we show that combining these findings yields a fast and robust adaptation algorithm for downstream tasks. Overall, with a single gradient step on 100 examples and a quick scan of the top candidate layers and factorization techniques, we can adapt LLMs to new datasets—entirely without fine‑tuning.

The rapid advancement of Large Language Models (LLMs) has catalyzed the development of multi-agent systems, where multiple LLM-based agents collaborate to solve complex tasks. However, existing systems predominantly rely on centralized coordination, which introduces scalability bottlenecks, limits adaptability, and creates single points of failure. Additionally, concerns over privacy and proprietary knowledge sharing hinder cross-organizational collaboration, leading to siloed expertise. To address these challenges, we propose AgentNet, a decentralized, Retrieval-Augmented Generation (RAG)-based framework that enables LLM-based agents to autonomously evolve their capabilities and collaborate efficiently in a Directed Acyclic Graph (DAG)-structured network. Unlike traditional multi-agent systems that depend on static role assignments or centralized control, AgentNet allows agents to specialize dynamically, adjust their connectivity, and route tasks without relying on predefined workflows. AgentNet’s core design is built upon several key innovations: (1) Fully Decentralized Paradigm: Removing the central orchestrator, allowing agents to coordinate and specialize autonomously, fostering fault tolerance and emergent collective intelligence. (2) Dynamically Evolving Graph Topology: Real-time adaptation of agent connections based on task demands, ensuring scalability and resilience. (3) Adaptive Learning for Expertise Refinement: A retrieval-based memory system that enables agents to continuously update and refine their specialized skills. By eliminating centralized control, AgentNet enhances fault tolerance, promotes scalable specialization, and enables privacy-preserving collaboration across organizations. Through decentralized coordination and minimal data exchange, agents can leverage diverse knowledge sources while safeguarding sensitive information. Experimental results demonstrate that AgentNet outperforms traditional centralized multi-agent systems, significantly improving efficiency, adaptability, and scalability in dynamic environments, making it a promising foundation for next-generation autonomous, privacy-respecting multi-agent ecosystems.

Watermarking LLM-generated text is critical for content attribution and misinformation prevention, yet existing methods compromise text quality and require white-box model access with logit manipulation or training, which exclude API-based models and multilingual scenarios. We propose SAEMark, an **inference-time framework** for *multi-bit* watermarking that embeds personalized information through *feature-based rejection sampling*, fundamentally different from logit-based or rewriting-based approaches: we **do not modify model outputs directly** and require only **black-box access**, while naturally supporting multi-bit message embedding and generalizing across diverse languages and domains. We instantiate the framework using *Sparse Autoencoders* as deterministic feature extractors and provide theoretical worst-case analysis relating watermark accuracy to computational budget. Experiments across 4 datasets demonstrate strong watermarking performance on English, Chinese, and code while preserving text quality. SAEMark establishes a new paradigm for **scalable, quality-preserving watermarks** that work seamlessly with closed-source LLMs across languages and domains.

Effective conversational agents must personalize their interactions to adapt to user preferences, personalities, and attributes across diverse domains like education and healthcare. Current methods like Reinforcement Learning from Human Feedback (RLHF), often prioritize helpfulness and safety but fall short in fostering truly empathetic, adaptive, and personalized dialogues. Existing personalization approaches typically rely on extensive user history, limiting their effectiveness for new or context-limited users. To address these limitations, we propose leveraging a user model to incorporate a curiosity-based intrinsic reward into multi-turn RLHF. This novel reward mechanism encourages the agent to actively infer user traits by optimizing conversations to improve its user model's accuracy. Consequently, the agent delivers more personalized interactions by learning more about the user. We demonstrate our method's effectiveness in two distinct domains: significantly improving personalization performance in a conversational recommendation task, and personalizing conversations for different learning styles in an educational setting with improved generalization capabilities compared to traditional multi-turn RLHF, all while maintaining conversation quality. Our method offers a promising solution for creating more personalized, adaptive, and engaging conversational agents.

Recent advances in Large Language Models (LLMs) have propelled intelligent agents from reactive responses to proactive support. While promising, existing proactive agents either rely exclusively on observations from enclosed environments (e.g., desktop UIs) with direct LLM inference or employ rule-based proactive notifications, leading to suboptimal user intent understanding and limited functionality for proactive service. In this paper, we introduce ContextAgent, the first context-aware proactive agent that incorporates extensive sensory contexts surrounding humans to enhance the proactivity of LLM agents. ContextAgent first extracts multi-dimensional contexts from massive sensory perceptions on wearables (e.g., video and audio) to understand user intentions. ContextAgent then leverages the sensory contexts and personas from historical data to predict the necessity for proactive services. When proactive assistance is needed, ContextAgent further automatically calls the necessary tools to assist users unobtrusively. To evaluate this new task, we curate ContextAgentBench, the first benchmark for evaluating context-aware proactive LLM agents, covering 1,000 samples across nine daily scenarios and twenty tools. Experiments on ContextAgentBench show that ContextAgent outperforms baselines by achieving up to 8.5% and 6.0% higher accuracy in proactive predictions and tool calling, respectively. We hope our research can inspire the development of more advanced, human-centric, proactive AI assistants. The code and dataset are publicly available at https://github.com/openaiotlab/ContextAgent.

We introduce **MLRC-Bench**, a benchmark designed to quantify how effectively language agents can tackle challenging **M**achine **L**earning (ML) **R**esearch **C**ompetitions, with a focus on open research problems that demand novel methodologies. Unlike prior work, e.g., AI Scientist, which evaluates the end-to-end agentic pipeline by using LLM-as-a-judge, MLRC-Bench measures the key steps of proposing and implementing novel research methods and evaluates them with rigorous protocol and objective metrics. Our curated suite of 7 competition tasks reveals significant challenges for LLM agents. Even the best-performing tested agent (gemini-exp-1206 under MLAB) closes only 9.3% of the gap between baseline and top human participant scores. Furthermore, our analysis reveals a misalignment between the *LLM-judged* innovation and their *actual* performance on cutting-edge ML research problems. MLRC-Bench is a dynamic benchmark, which is designed to continually grow with new ML competitions to encourage rigorous and objective evaluations of AI’s research capabilities. Our leaderboard and code are publicly available at https://huggingface.co/spaces/launch/MLRC_Bench.

Large Language Models exhibit logical inconsistency across multi-turn inference processes, undermining correctness in complex inferential tasks. Challenges arise from ensuring that outputs align with both factual correctness and human intent. Approaches like single-agent reflection and multi-agent debate frequently prioritize consistency, but at the expense of accuracy. To address this problem, we propose a novel game-theoretic consensus mechanism that enables LLMs to self-check their outputs during the decoding stage of output generation. Our method models the decoding process as a multistage Bayesian Decoding Game, where strategic interactions dynamically converge to a consensus on the most reliable outputs without human feedback or additional training. Remarkably, our game design allows smaller models to outperform much larger models through game mechanisms (e.g., 78.1 LLaMA13B vs. 76.6 PaLM540B). As a model-agnostic method, our approach consistently improves even the latest models, enhancing DeepSeek-7B's performance on MMLU by 12.4%. Our framework effectively balances correctness and consistency, demonstrating that properly designed game-theoretic mechanisms can significantly enhance the self-verification capabilities of language models across various tasks and model architectures.

In cooperative Multi-Agent Reinforcement Learning (MARL), it is a common practice to tune hyperparameters in ideal simulated environments to maximize cooperative performance. However, policies tuned for cooperation often fail to maintain robustness and resilience under real-world uncertainties. Building trustworthy MARL systems requires a deep understanding of \emph{robustness}, which ensures stability under uncertainties, and \emph{resilience}, the ability to recover from disruptions—a concept extensively studied in control systems but largely overlooked in MARL. In this paper, we present a large-scale empirical study comprising over 82,620 experiments to evaluate cooperation, robustness, and resilience in MARL across 4 real-world environments, 13 uncertainty types, and 15 hyperparameters. Our key findings are: (1) Under mild uncertainty, optimizing cooperation improves robustness and resilience, but this link weakens as perturbations intensify. Robustness and resilience also varies by algorithm and uncertainty type. (2) Robustness and resilience do not generalize across uncertainty modalities or agent scopes: policies robust to action noise for all agents may fail under observation noise on a single agent. (3) Hyperparameter tuning is critical for trustworthy MARL: surprisingly, standard practices like parameter sharing, GAE, and PopArt can hurt robustness, while early stopping, high critic learning rates, and Leaky ReLU consistently help. By optimizing hyperparameters only, we observe substantial improvement in cooperation, robustness and resilience across all MARL backbones, with the phenomenon also generalizing to robust MARL methods across these backbones.

3D generation from natural language offers significant potential to reduce expert manual modeling efforts and enhance accessibility to 3D assets. However, existing methods often yield unstructured meshes and exhibit poor interactivity, making them impractical for artistic workflows. To address these limitations, we represent 3D assets as shape programs and introduce ShapeCraft, a novel multi-agent framework for text-to-3D generation. At its core, we propose a Graph-based Procedural Shape (GPS) representation that decomposes complex natural language into a structured graph of sub-tasks, thereby facilitating accurate LLM comprehension and interpretation of spatial relationships and semantic shape details. Specifically, LLM agents hierarchically parse user input to initialize GPS, then iteratively refine procedural modeling and painting to produce structured, textured, and interactive 3D assets. Qualitative and quantitative experiments demonstrate ShapeCraft's superior performance in generating geometrically accurate and semantically rich 3D assets compared to existing LLM-based agents. We further show the versatility of ShapeCraft through examples of animated and user-customized editing, highlighting its potential for broader interactive applications.

Ensembles of generative large language models (LLMs) are a promising way to compensate for individual model limitations, integrating the strengths of different LLMs. Existing LLM ensemble methods, however, face limitations such as first-token delay and challenges in long-range semantic collaboration between models, Moreover, they typically assume equal voting weights for all models during ensemble, ignoring performance differences between models for a given task. In this work, we propose SpecEM, a training-free, plug-and-play LLM ensemble framework that dynamically adjusts each model's model contribution in real time based on task performance. Inspired by speculative decoding, SpecFuse iteratively performs drafting and verification, allowing models to collaborate semantically at the segment level for integrated output. Furthermore, we introduce an online feedback mechanism with multiplicative weight updates, where each model's voting weight is adjusted on-the-fly according to how often it "outperforms" others during verification stage, ensuring that stronger models exert greater influence on the ensemble during generation. Experimental results on five popular LLMs (ranging from 7B to 72B parameters) and six benchmark tasks, spanning instruction following, reasoning, commonsense, and general instruction response, demonstrate consistent performance improvements compared to state-of-the-art LLM ensemble methods.

Large language models (LLMs) have been widely integrated into information retrieval to advance traditional techniques. However, effectively enabling LLMs to seek accurate knowledge in complex tasks remains a challenge due to the complexity of multi-hop queries as well as the irrelevant retrieved content. To address these limitations, we propose ExSearch, an agentic search framework, where the LLM learns to retrieve useful information as the reasoning unfolds through a self-incentivized process. At each step, the LLM decides what to retrieve (thinking), triggers an external retriever (search), and extracts fine-grained evidence (recording) to support next-step reasoning. To enable LLM with this capability, we adopts a Generalized Expectation-Maximization algorithm. In the E-step, the LLM generates multiple search trajectories and assigns an importance weight to each; the M-step trains the LLM on them with a re-weighted loss function. This creates a self-incentivized loop, where the LLM iteratively learns from its own generated data, progressively improving itself for search. We further theoretically analyze this training process, establishing convergence guarantees. Extensive experiments on four knowledge-intensive benchmarks show that ExSearchS substantially outperforms baselines, e.g., +7.8% improvement on exact match score. Motivated by these promising results, we introduce ExSearch-Zoo, an extension that extends our method to broader scenarios, to facilitate future work.

Conventional research on large language models (LLMs) has primarily focused on refining output distributions, while paying less attention to the decoding process that transforms these distributions into final responses. Recent advances, such as scaling the computation of inference time with reward models, have underscored the importance of decoding, but these methods often suffer from high computational costs and limited applicability. In this paper, we revisit LLM generation through the lens of recommender systems, conceptualizing the decoding process as analogous to the ranking stage in recommendation pipelines. From this perspective, we observe that both traditional decoding methods and reward models exhibit clear limitations such as redundancy. Motivated by this insight, we propose Language Ranker, a novel framework that introduces a lightweight module to rerank candidate responses using features extracted by the base model. Experiments across a wide range of tasks show that Language Ranker achieves performance comparable to large-scale reward models, while requiring only <0.5M additional parameters, significantly reducing the computational overhead during both training and inference stages. This highlights the efficiency and effectiveness of our method, showcasing its potential to fully unlock the capabilities of LLMs.

Human social interactions depend on the ability to infer others' unspoken intentions, emotions, and beliefs—a cognitive skill grounded in the psychological concept of Theory of Mind (ToM). While large language models (LLMs) excel in semantic understanding tasks, they struggle with the ambiguity and contextual nuance inherent in human communication. To bridge this gap, we introduce **MetaMind**, a multi-agent framework inspired by psychological theories of metacognition, designed to emulate human-like social reasoning. MetaMind decomposes social understanding into three collaborative stages: (1) a *Theory-of-Mind Agent* generates hypotheses about user mental states (e.g., intent, emotion), (2) a *Moral Agent* refines these hypotheses using cultural norms and ethical constraints, and (3) a *Response Agent* generates contextually appropriate responses while validating alignment with inferred intent. Our framework achieves state-of-the-art performance across three challenging benchmarks, with 35.7% improvement in real-world social scenarios and 6.2% gain in ToM reasoning. Notably, it enables LLMs to match human-level performance on key ToM tasks for the first time. Ablation studies confirm the necessity of all components, which showcase the framework’s ability to balance contextual plausibility, social appropriateness, and user adaptation. This work advances AI systems toward human-like social intelligence, with applications in empathetic dialogue and culturally sensitive interactions. Code is available at https://github.com/XMZhangAI/MetaMind.

LLM-agent systems often decompose high-level objectives into subtask dependency graphs, assuming that each subtask’s output is reliable and conditionally independent of others given its parent responses. However, this assumption frequently breaks during execution, as ground-truth responses are inaccessible, leading to inter-agent misalignment—failures caused by inconsistencies and coordination breakdowns among agents. To address this, we propose SeqCV, a dynamic framework for reliable execution under violated conditional independence. SeqCV executes subtasks sequentially, each conditioned on all prior verified responses, and performs consistency checks immediately after agents generate short token sequences. At each checkpoint, a token sequence is accepted only if it represents shared knowledge consistently supported across diverse LLM models; otherwise, it is discarded, triggering recursive subtask decomposition for finer-grained reasoning. Despite its sequential nature, SeqCV avoids repeated corrections on the same misalignment and achieves higher effective throughput than parallel pipelines. Across multiple reasoning and coordination tasks, SeqCV improves accuracy by up to 30\% over existing LLM-agent systems. Code is available at https://github.com/tmllab/2025_NeurIPS_SeqCV.

Long-horizon tasks requiring multi-step reasoning and dynamic re-planning remain challenging for large language models (LLMs). Sequential prompting methods are prone to context drift, loss of goal information, and recurrent failure cycles, while hierarchical prompting methods often weaken cross-level continuity or incur substantial runtime overhead. We introduce ReCAP (Recursive Context-Aware Reasoning and Planning), a hierarchical framework with shared context for reasoning and planning in LLMs. ReCAP combines three key mechanisms: (i) plan-ahead decomposition, in which the model generates a full subtask list, executes the first item, and refines the remainder; (ii) structured re-injection of parent plans, maintaining consistent multi-level context during recursive return; and (iii) memory-efficient execution, bounding the active prompt so costs scale linearly with task depth. Together these mechanisms align high-level goals with low-level actions, reduce redundant prompting, and preserve coherent context updates across recursion. Experiments demonstrate that ReCAP substantially improves subgoal alignment and success rates on various long-horizon reasoning benchmarks, achieving a 32\% gain on synchronous Robotouille and a 29\% improvement on asynchronous Robotouille under the strict pass@1 protocol.

Multimodal Large Language Models (MLLMs) have shown great potential in revolutionizing Graphical User Interface (GUI) automation. However, existing GUI models mostly rely on learning from nearly error-free offline trajectories, thus lacking reflection and error recovery capabilities. To bridge this gap, we propose GUI-Reflection, a novel framework that explicitly integrates self-reflection and error correction capabilities into end-to-end multimodal GUI models throughout dedicated training stages: GUI-specific pre-training, offline supervised fine-tuning (SFT), and online reflection tuning. GUI-reflection enables self-reflection behavior emergence with fully automated data generation and learning processes without requiring any human annotation. Specifically, 1) we first propose scalable data pipelines to automatically construct reflection and error correction data from existing successful trajectories. While existing GUI models mainly focus on grounding and UI understanding ability, we propose the GUI-Reflection Task Suite to learn and evaluate reflection-oriented abilities explicitly. 2) Furthermore, we built a diverse and efficient environment for online training and data collection of GUI models on mobile devices. 3) We also present an iterative online reflection tuning algorithm leveraging the proposed environment, enabling the model to continuously enhance its reflection and error correction abilities. Our framework equips GUI agents with self-reflection and correction capabilities, paving the way for more robust, adaptable, and intelligent GUI automation, with all data, models, environments, and tools to be released publicly.

Addressing intricate real-world problems necessitates in-depth information seeking and multi-step reasoning. Recent progress in agentic systems, exemplified by Deep Research, underscores the potential for autonomous multi-step research. In this work, we present a cohesive paradigm for building end-to-end agentic information seeking agents from a data-centric and training-stage perspective. Our approach consists of four key stages: (1) browsing data construction, (2) trajectories sampling, (3) supervised fine-tuning for effective cold start, and (4) reinforcement learning for enhanced generalisation. We instantiate this framework in a web agent based on the ReAct format, WebDancer. Empirical evaluations on the challenging GAIA and WebWalkerQA benchmarks demonstrate the strong performance of WebDancer, achieving considerable results and highlighting the efficacy of our training paradigm. Further analysis of agent training provides valuable insights and actionable, systematic pathways for developing more capable agentic models.