NeurIPS 2025 论文列表

We propose a new algorithm for fine-tuning large language models using reinforcement learning. Tapered Off-Policy REINFORCE (TOPR) uses an asymmetric, tapered variant of importance sampling to speed up learning while maintaining stable learning dynamics, even without the use of KL regularization. TOPR can be applied in a fully offline fashion, allows the handling of positive and negative examples in a unified framework, and benefits from the implementational simplicity that is typical of Monte Carlo algorithms. We demonstrate the effectiveness of our approach with a series of experiments on the GSM8K and MATH reasoning benchmarks, finding performance gains for training both a model for solution generation and as a generative verifier. We show that properly leveraging positive and negative examples alike in the off-policy regime simultaneously increases test-time accuracy and training data efficiency, all the while avoiding the ``wasted inference'' that comes with discarding negative examples. We find that this advantage persists over multiple iterations of training and can be amplified by dataset curation techniques, enabling us to match 70B-parameter model performance with 8B language models. As a corollary to this work, we find that REINFORCE's baseline parameter plays an important and unexpected role in defining dataset composition in the presence of negative examples, and is consequently critical in driving off-policy performance.

Chip placement is a critical step in the Electronic Design Automation (EDA) workflow, which aims to arrange chip modules on the canvas to optimize the performance, power, and area (PPA) metrics of final designs. Recent advances show great potential of AI-based algorithms in chip placement. However, due to the lengthy EDA workflow, evaluations of these algorithms often focus on intermediate surrogate metrics, which are computationally efficient but often misalign with the final end-to-end performance (i.e., the final design PPA). To address this challenge, we propose to build ChiPBench, a comprehensive benchmark specifically designed to evaluate the effectiveness of AI-based algorithms in final design PPA metrics. Specifically, we generate a diverse evaluation dataset from $20$ circuits across various domains, such as CPUs, GPUs, and NPUs. We then evaluate six state-of-the-art AI-based chip placement algorithms on the dataset and conduct a thorough analysis of their placement behavior. Extensive experiments show that AI-based chip placement algorithms produce unsatisfactory final PPA results, highlighting the significant influence of often-overlooked factors like regularity and dataflow. We believe ChiPBench will effectively bridge the gap between academia and industry.

Large Language Models (LLMs) show great promise in complex reasoning, with Reinforcement Learning with Verifiable Rewards (RLVR) being a key enhancement strategy. However, a prevalent issue is ``superficial self-reflection'', where models fail to robustly verify their own outputs. We introduce RISE (Reinforcing Reasoning with Self-Verification), a novel online RL framework designed to tackle this. RISE explicitly and simultaneously trains an LLM to improve both its problem-solving and self-verification abilities within a single, integrated RL process. The core mechanism involves leveraging verifiable rewards from an outcome verifier to provide on-the-fly feedback for both solution generation and self-verification tasks. In each iteration, the model generates solutions, then critiques its own on-policy generated solutions, with both trajectories contributing to the policy update. Extensive experiments on diverse mathematical reasoning benchmarks show that RISE consistently improves model's problem-solving accuracy while concurrently fostering strong self-verification skills. Our analyses highlight the advantages of online verification and the benefits of increased verification compute. Additionally, RISE models exhibit more frequent and accurate self-verification behaviors during reasoning. These advantages reinforce RISE as a flexible and effective path towards developing more robust and self-aware reasoners.

Large language models (LLMs) have achieved remarkable achievements across diverse applications; however, they remain plagued by spurious correlations and the generation of hallucinated content. Despite extensive efforts to enhance the resilience of LLMs, existing approaches either rely on indiscriminate fine-tuning of all parameters, resulting in parameter inefficiency and lack of specificity, or depend on post-processing techniques that offer limited adaptability and flexibility. This study introduces a novel Causality-driven Robust Optimization (CdRO) approach that selectively updates model components sensitive to causal reasoning, enhancing model causality while preserving valuable pretrained knowledge to mitigate overfitting. Our method begins by identifying the parameter components within LLMs that capture causal relationships, achieved through comparing the training dynamics of parameter matrices associated with the original samples, as well as augmented counterfactual and paraphrased variants. These comparisons are then fed into a lightweight logistic regression model, optimized in real time to dynamically identify and adapt the causal components within LLMs. The identified parameters are subsequently optimized using an enhanced policy optimization algorithm, where the reward function is designed to jointly promote both model generalization and robustness. Extensive experiments across various tasks using twelve different LLMs demonstrate the superior performance of our framework, underscoring its significant effectiveness in reducing the model’s dependence on spurious associations and mitigating hallucinations.

Autonomy is a hallmark of animal intelligence, enabling adaptive and intelligent behavior in complex environments without relying on external reward or task structure. Existing reinforcement learning approaches to exploration in reward-free environments, including a class of methods known as *model-based intrinsic motivation*, exhibit inconsistent exploration patterns and do not converge to an exploratory policy, thus failing to capture robust autonomous behaviors observed in animals. Moreover, systems neuroscience has largely overlooked the neural basis of autonomy, focusing instead on experimental paradigms where animals are motivated by external reward rather than engaging in ethological, naturalistic and task-independent behavior. To bridge these gaps, we introduce a novel model-based intrinsic drive explicitly designed after the principles of autonomous exploration in animals. Our method (3M-Progress) achieves animal-like exploration by tracking divergence between an online world model and a fixed prior learned from an ecological niche. To the best of our knowledge, we introduce the first autonomous embodied agent that predicts brain data entirely from self-supervised optimization of an intrinsic goal—without any behavioral or neural training data—demonstrating that 3M-Progress agents capture the explainable variance in behavioral patterns and whole-brain neural-glial dynamics recorded from autonomously behaving larval zebrafish, thereby providing the first goal-driven, population-level model of neural-glial computation. Our findings establish a computational framework connecting model-based intrinsic motivation to naturalistic behavior, providing a foundation for building artificial agents with animal-like autonomy.

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for training language models (LMs) on reasoning tasks that elicit emergent long chains of thought (CoTs). Unlike supervised learning, it updates the model using both correct and incorrect samples via policy gradients. To better understand its mechanism, we decompose the learning signal into reinforcing correct responses and penalizing incorrect ones, referred to as **P**ositive and **N**egative **S**ample **R**einforcement (**PSR** and **NSR**), respectively. We train `Qwen2.5-Math-7B`, `Qwen3-4B` and `Llama-3.1-8B-Instruct` on a mathematical reasoning dataset and uncover a surprising result: training with only negative samples — without reinforcing correct responses — can be highly effective: it consistently improves performance over the base model across the entire Pass@$k$ spectrum $k$ up to 256), often matching or surpassing PPO and GRPO. In contrast, reinforcing only correct responses improves Pass@1 but degrades performance at higher $k$, due to reduced diversity. These inference-scaling trends highlight that solely penalizing incorrect responses may contribute more to performance than previously recognized. Through gradient analysis, we show that NSR works by suppressing incorrect generations and redistributing probability mass toward other plausible candidates, guided by the model's prior beliefs. It refines the model's existing knowledge rather than introducing entirely new behaviors. Building on this insight, we propose a simple variant of the RL objective that upweights NSR, and show that it consistently improves overall Pass@$k$ performance on MATH, AIME 2025, and AMC23. Our code is available at [`https://github.com/TianHongZXY/RLVR-Decomposed`](https://github.com/TianHongZXY/RLVR-Decomposed).

The iterative and stochastic nature of diffusion models enables *test-time scaling*, whereby spending additional compute during denoising generates higher-fidelity samples. Increasing the number of denoising steps is the primary scaling axis, but this yields quickly diminishing returns. Instead optimizing the *noise trajectory*—the sequence of injected noise vectors—is promising, as the specific noise realizations critically affect sample quality; but this is challenging due to a high-dimensional search space, complex noise-outcome interactions, and costly trajectory evaluations. We address this by first casting diffusion as a Markov Decision Process (MDP) with a terminal reward, showing tree-search methods such as Monte Carlo tree search (MCTS) to be meaningful but impractical. To balance performance and efficiency, we then resort to a relaxation of MDP, where we view denoising as a sequence of independent *contextual bandits*. This allows us to introduce an $\epsilon$-greedy search algorithm that *globally explores* at extreme timesteps and *locally exploits* during the intermediate steps where de-mixing occurs. Experiments on EDM and Stable Diffusion reveal state-of-the-art scores for class-conditioned/text-to-image generation, exceeding baselines by up to $164$% and matching/exceeding MCTS performance. To our knowledge, this is the first practical method for test-time noise *trajectory* optimization of *arbitrary (non-differentiable)* rewards.

Large Language Models (LLMs), such as OpenAI’s o1 and DeepSeek’s R1, excel at advanced reasoning tasks like math and coding via Reinforcement Learning with Verifiable Rewards (RLVR), but still struggle with puzzles solvable by humans without domain knowledge. We introduce ENIGMATA, the first comprehensive suite tailored for improving LLMs with puzzle reasoning skills. It includes 36 tasks across 7 categories, each with: 1) a generator that produces unlimited examples with controllable difficulty, and 2) a rule-based verifier for automatic evaluation. This generator-verifier design supports scalable, multi-task RL training, fine-grained analysis, and seamless RLVR integration. We further propose ENIGMATA-Eval, a rigorous benchmark, and develop optimized multi-task RLVR strategies. Our trained model, Qwen2.5-32B-ENIGMATA, consistently surpasses o3-mini-high and o1 on the puzzle reasoning benchmarks like ENIGMATA-Eval, ARC-AGI (32.8%), and ARC-AGI 2 (0.6%). It also generalizes well to out-of-domain puzzle benchmarks and mathematical reasoning, with little multi-tasking trade-off. When trained on larger models like Seed1.5-Thinking (20B activated parameters and 200B total parameters), puzzle data from ENIGMATA further boosts SoTA performance on advanced math and STEM reasoning tasks such as AIME (2024-2025), BeyondAIME and GPQA (Diamond), showing nice generalization benefits of ENIGMATA. This work offers a unified, controllable framework for advancing logical reasoning in LLMs. Project page: https://seed-enigmata.github.io.

Multilingual reasoning remains a significant challenge for large language models (LLMs), with performance disproportionately favoring high-resource languages. Drawing inspiration from cognitive neuroscience, which suggests that human reasoning functions largely independently of language processing, we hypothesize that LLMs similarly encode reasoning and language as separable components that can be disentangled to enhance multilingual reasoning. To evaluate this, we perform a causal intervention by ablating language-specific representations at inference time. Experiments on 10 open-weight LLMs spanning 11 typologically diverse languages show that this language-specific ablation consistently boosts multilingual reasoning performance. Layer-wise analyses further confirm that language and reasoning representations can be effectively disentangled throughout the model, yielding improved multilingual reasoning capabilities, while preserving top-layer language features remains essential for maintaining linguistic fidelity. Compared to post-training methods such as supervised fine-tuning or reinforcement learning, our training-free language-reasoning disentanglement achieves comparable or superior results with minimal computational overhead. These findings shed light on the internal mechanisms underlying multilingual reasoning in LLMs and suggest a lightweight and interpretable strategy for improving cross-lingual generalization.

Multi-task multi-agent reinforcement learning (MT-MARL) aims to develop a single model capable of solving a diverse set of tasks. However, existing methods often fall short due to the substantial variation in optimal policies across tasks, making it challenging for a single policy model to generalize effectively. In contrast, we find that many tasks exhibit **bounded similarity** in their underlying dynamics—highly similar within certain groups (e.g., door-open/close) diverge significantly between unrelated tasks (e.g., door-open \& object-catch). To leverage this property, we reconsider the role of modularity in multi-task learning, and propose **M3W**, a novel approach that applies mixture-of-experts (MoE) to world model instead of policy, enabling both learning and planning. For learning, it uses a SoftMoE-based dynamics model alongside a SparseMoE-based predictor to facilitate knowledge reuse across similar tasks while avoiding gradient conflicts across dissimilar tasks. For planning, it evaluates and optimizes actions using the predicted rollouts from the world model, without relying directly on a explicit policy model, thereby overcoming the limitations of policy-centric methods. As the first MoE-based multi-task world model, M3W demonstrates superior performance, sample efficiency, and multi-task adaptability, as validated on Bi-DexHands with 14 tasks and MA-Mujoco with 24 tasks. The demos and anonymous code are available at \url{https://github.com/zhaozijie2022/m3w-marl}.

Large Vision-Language Models (LVLMs) show promise for embodied planning tasks but struggle with complex scenarios involving unfamiliar environments and multi-step goals. Current approaches rely on environment-agnostic imitation learning that disconnects instructions from environmental contexts, causing models to struggle with context-sensitive instructions and rely on supplementary cues rather than visual reasoning during long-horizon interactions. In this work, we propose World-Aware Planning Narrative Enhancement (WAP), a framework that infuses LVLMs with comprehensive environmental understanding through four cognitive capabilities (visual appearance modeling, spatial reasoning, functional abstraction, and syntactic grounding) while developing and evaluating models using only raw visual observations through curriculum learning. Evaluations on the EB-ALFRED benchmark demonstrate substantial improvements, with Qwen2.5-VL achieving a 60.7 absolute improvement in task success rates—particularly in commonsense reasoning (+60.0) and long-horizon planning (+70.0). Notably, our enhanced open-source models outperform proprietary systems like GPT-4o and Claude-3.5-Sonnet by a large margin.

Recently, leveraging pre-trained vision-language models (VLMs) for building vision-language-action (VLA) models has emerged as a promising approach to effective robot manipulation learning. However, only few methods incorporate 3D signals into VLMs for action prediction, and they do not fully leverage the spatial structure inherent in 3D data, leading to low data efficiency. In this paper, we introduce a new paradigm for constructing 3D VLAs. Specifically, we first pre-train the VLM backbone to take 2D images as input and produce 2D heatmaps as output. Using this pre-trained VLM as the backbone, we then fine-tune the entire VLA model while maintaining alignment between inputs and outputs by: (1) projecting raw point cloud inputs into multi-view images, and (2) predicting heatmaps before generating the final action. Extensive experiments show that the resulting model, BridgeVLA, can learn 3D manipulation both efficiently and effectively. BridgeVLA outperforms state-of-the-art baselines across three simulation benchmarks. In RLBench, it improves the average success rate from 81.4\% to 88.2\%. In COLOSSEUM, it demonstrates significantly better performance in challenging generalization settings, boosting the average success rate from 56.7\% to 64.0\%. In GemBench, it surpasses all the comparing baseline methods in terms of average success rate. In real-robot experiments, BridgeVLA outperforms a state-of-the-art baseline method by 32\% on average. It generalizes robustly in multiple out-of-distribution settings, including visual disturbances and unseen instructions. Remarkably, it is able to achieve a success rate of 95.4\% on 10+ tasks with only 3 trajectories per task, while other VLA methods such as $\pi_{0}$ fail completely. Project Website: https://bridgevla.github.io/.

We address the problem of safe policy learning in multi-agent safety-critical autonomous systems. In such systems, it is necessary for each agent to meet the safety requirements at all times while also cooperating with other agents to accomplish the task. Toward this end, we propose a safe Hierarchical Multi-Agent Reinforcement Learning (HMARL) approach based on Control Barrier Functions (CBFs). Our proposed hierarchical approach decomposes the overall reinforcement learning problem into two levels –- learning joint cooperative behavior at the higher level and learning safe individual behavior at the lower or agent level conditioned on the high-level policy. Specifically, we propose a skill-based HMARL-CBF algorithm in which the higher-level problem involves learning a joint policy over the skills for all the agents and the lower-level problem involves learning policies to execute the skills safely with CBFs. We validate our approach on challenging environment scenarios whereby a large number of agents have to safely navigate through conflicting road networks. Compared with existing state-of-the-art methods, our approach significantly improves the safety achieving near perfect (within $5\%$) success/safety rate while also improving performance across all the environments.

The field of embodied AI (EAI) is rapidly advancing. Unlike virtual AI, EAI systems can exist in, learn from, reason about, and act in the physical world. With recent advances in AI and hardware research and design, EAI systems are becoming increasingly capable across an expanding set of operational domains. While EAI systems can offer many benefits, they also pose significant short- and long-term risks, including physical harm, surveillance, and societal disruption. These risks require urgent attention from policymakers, as existing policies for industrial robots and autonomous vehicles are insufficient to manage the full range of concerns EAI systems present. To address this issue, this paper makes three contributions. First, we provide a taxonomy of the physical, informational, economic, and social risks EAI systems pose. Second, we analyze policies in the US, UK, and EU to assess how existing frameworks address these risks and to identify critical gaps. We conclude by offering policy recommendations for the safe and beneficial deployment of EAI systems, such as mandatory testing and certification schemes, clarified liability frameworks, and strategies to manage EAI’s potentially transformative economic and societal impacts.

Offline reinforcement learning (offline RL) is increasingly approached as a sequence modeling task, with methods leveraging advanced architectures like Transformers to capture trajectory dependencies. Despite significant progress, the mechanisms underlying their effectiveness and limitations remain insufficiently understood. We conduct a thorough analysis on the representative Decision Transformer (DT) model using an entropy analysis and identify the inconsistencies in state-action-reward ($\langle s, a, R \rangle$) distributions causing attention ``dispersal". To address this, we propose a hierarchical framework that decomposes sequence modeling into intra-step relational modeling—handled by a Token Merger that fuses each $\langle s, a, R \rangle$ triplet—and inter-step modeling—handled by a Token Mixer across timesteps. We investigate several Token Merger designs and validate their effectiveness across various offline RL methods. Furthermore, our theoretical analysis and experimental results suggest that while Token Mixers are important, lightweight architecture can also achieve even better performance to more complex ones. We therefore propose a parameter-free Average Pooling Token Mixer, which, combined with a convolutional Token Merger, forms our final model, Decision HiFormer (DHi). DHi achieves a \textbf{73.6\%} improvement in inference speed and an \textbf{9.3\%} gain in policy performance on the D4RL benchmark compared to DT. DHi also generalizes well to real-world robotic manipulation tasks, offering both practical benefits and insights into sequence-based policy design for offline RL. Code and models are public at \href{https://wei-nijuan.github.io/DecisionHiFormer/}{project page}.

The core task of recommender systems is to learn user preferences from historical user-item interactions. With the rapid development of large language models (LLMs), recent research has explored leveraging the reasoning capabilities of LLMs to enhance rating prediction tasks. However, existing distillation-based methods suffer from limitations such as the teacher model's insufficient recommendation capability, costly and static supervision, and superficial transfer of reasoning ability. To address these issues, this paper proposes RecZero, a reinforcement learning (RL)-based recommendation paradigm that abandons the traditional multi-model and multi-stage distillation approach. Instead, RecZero trains a single LLM through pure RL to autonomously develop reasoning capabilities for rating prediction. RecZero consists of two key components: (1) "Think-before-Recommendation" prompt construction, which employs a structured reasoning template to guide the model in step-wise analysis of user interests, item features, and user-item compatibility; and (2) rule-based reward modeling, which adopts group relative policy optimization (GRPO) to compute rewards for reasoning trajectories and optimize the LLM. Additionally, the paper explores a hybrid paradigm, RecOne, which combines supervised fine-tuning with RL, initializing the model with cold-start reasoning samples and further optimizing it with RL. Experimental results demonstrate that RecZero and RecOne significantly outperform existing baseline methods on multiple benchmark datasets, validating the superiority of the RL paradigm in achieving autonomous reasoning-enhanced recommender systems.

Motivated by real-world settings where data collection and policy deployment—whether for a single agent or across multiple agents—are costly, we study the problem of on-policy single-agent reinforcement learning (RL) and federated RL (FRL) with a focus on minimizing burn-in costs (the sample sizes needed to reach near-optimal regret) and policy switching or communication costs. In parallel finite-horizon episodic Markov Decision Processes (MDPs) with $S$ states and $A$ actions, existing methods either require superlinear burn-in costs in $S$ and $A$ or fail to achieve logarithmic switching or communication costs. We propose two novel model-free RL algorithms—Q-EarlySettled-LowCost and FedQ-EarlySettled-LowCost—that are the first in the literature to simultaneously achieve: (i) the best near-optimal regret among all known model-free RL or FRL algorithms, (ii) low burn-in cost that scales linearly with $S$ and $A$, and (iii) logarithmic policy switching cost for single-agent RL or communication cost for FRL. Additionally, we establish gap-dependent theoretical guarantees for both regret and switching/communication costs, improving or matching the best-known gap-dependent bounds.

The success of Direct Preference Optimization (DPO) in mitigating hallucinations in Vision Language Models (VLMs) critically hinges on the true reward gaps within preference pairs. However, current methods, typically relying on ranking or rewriting strategies, often struggle to optimize these reward gaps in a systematic way during data curation. A core difficulty lies in precisely characterizing and strategically manipulating the overall reward gap configuration, that is, the deliberate design of how to shape these reward gaps within each preference pair across the data. To address this, we introduce Topic-level Preference Rewriting (TPR), a novel framework designed for the systematic optimization of reward gap configuration. Through selectively replacing semantic topics within VLM responses with model’s own resampled candidates for targeted rewriting, TPR can provide topic-level control over fine-grained semantic details. This precise control enables advanced data curation strategies, such as progressively adjusting the difficulty of rejected responses, thereby sculpting an effective reward gap configuration that guides the model to overcome challenging hallucinations. Comprehensive experiments demonstrate TPR achieves state-of-the-art performance on multiple hallucination benchmarks, outperforming previous methods by an average of $\sim$20%. Notably, it significantly reduces hallucinations by up to 93% on ObjectHal-Bench, and also exhibits superior data efficiency towards robust and cost-effective VLM alignment.

Recent Large Reasoning Models (LRMs) have shown substantially improved reasoning capabilities over traditional Large Language Models (LLMs) by incorporating extended thinking processes prior to producing final responses. However, excessively lengthy thinking introduces substantial overhead in terms of token consumption and latency, which is unnecessary for simple queries. In this work, we introduce Large Hybrid-Reasoning Models (LHRMs), the first kind of model capable of adaptively determining whether to perform reasoning based on the contextual information of user queries. To achieve this, we propose a two-stage training pipeline comprising Hybrid Fine-Tuning (HFT) as a cold start, followed by online reinforcement learning with the proposed Hybrid Group Policy Optimization (HGPO) to implicitly learn to select the appropriate reasoning mode. Furthermore, we introduce a metric called Hybrid Accuracy to quantitatively assess the model’s capability for hybrid reasoning. Extensive experimental results show that LHRMs can adaptively perform hybrid reasoning on queries of varying difficulty and type. It outperforms existing LRMs and LLMs in reasoning and general capabilities while significantly improving efficiency. Together, our work advocates for a reconsideration of the appropriate use of extended reasoning processes and provides a solid starting point for building hybrid reasoning systems.