ICML 2025 论文列表

In this paper, we present DiffusionVLA, a novel framework that integrates autoregressive reasoning with diffusion policies to address the limitations of existing methods: while autoregressive Vision-Language-Action (VLA) models lack precise and robust action generation, diffusion-based policies inherently lack reasoning capabilities. Central to our approach is autoregressive reasoning — a task decomposition and explanation process enabled by a pre-trained VLM — to guide diffusion-based action policies. To tightly couple reasoning with action generation, we introduce a reasoning injection module that directly embeds self-generated reasoning phrases into the policy learning process. The framework is simple, flexible, and efficient, enabling seamless deployment across diverse robotic platforms. We conduct extensive experiments using multiple real robots to validate the effectiveness of DiVLA. Our tests include a challenging factory sorting task, where DiVLA successfully categorizes objects, including those not seen during training. The reasoning injection module enhances interpretability, enabling explicit failure diagnosis by visualizing the model’s decision process. Additionally, we test DiVLA on a zero-shot bin-picking task, achieving \textbf{63.7\% accuracy on 102 previously unseen objects}. Our method demonstrates robustness to visual changes, such as distractors and new backgrounds, and easily adapts to new embodiments. Furthermore, DiVLA can follow novel instructions and retain conversational ability. Notably, DiVLA is data-efficient and fast at inference; our smallest DiVLA-2B runs 82Hz on a single A6000 GPU. Finally, we scale the model from 2B to 72B parameters, showcasing improved generalization capabilities with increased model size.

While language models have exceptional capabilities at text generation, they lack a natural inductive bias for emitting numbers and thus struggle in tasks involving quantitative reasoning, especially arithmetic. One fundamental limitation is the nature of the Cross Entropy loss, which assumes a nominal scale and thus cannot convey proximity between generated number tokens. In response, we here present a regression-like loss that operates purely on token level. Our proposed **Number Token Loss** (NTL) comes in two flavors and minimizes either the $\mathcal{L}_p$ norm or the Wasserstein distance between the *numerical values* of the real and predicted number tokens. NTL can easily be added to any language model and extend the Cross Entropy objective during training without runtime overhead. We evaluate the proposed scheme on various mathematical datasets and find that it consistently improves performance in math-related tasks. In a direct comparison on a regression task, we find that NTL can match the performance of a regression head, despite operating on token level. Finally, we scale NTL up to 3B parameter models and observe improved performance, demonstrating its potential for seamless integration into LLMs. We hope that this work can inspire LLM developers to improve their pretraining objectives. The code is available via: https://tum-ai.github.io/number-token-loss/

Chain-of-Thought (CoT) technique has proven effective in improving the performance of large language models (LLMs) on complex reasoning tasks. However, the performance gains are inconsistent across different tasks, and the underlying mechanism remains a long-standing research question. In this work, we make a preliminary observation that the monotonicity of token probability distributions may be correlated with the gains achieved through CoT reasoning. Leveraging this insight, we propose two indicators based on the token probability distribution to assess CoT effectiveness across different tasks. By combining instance-level indicators with logistic regression model, we introduce Dynamic CoT, a method that dynamically select between CoT and direct answer. Furthermore, we extend Dynamic CoT to closed-source models by transferring decision strategies learned from open-source models. Our indicators for assessing CoT effectiveness achieve an accuracy of 89.2\%, and Dynamic CoT reduces token consumption by more than 35\% while maintaining high accuracy. Overall, our work offers a novel perspective on the underlying mechanisms of CoT reasoning and provides a framework for its more efficient deployment.

We propose Noise Conditional Variational Score Distillation (NCVSD), a novel method for distilling pretrained diffusion models into generative denoisers. We achieve this by revealing that the unconditional score function implicitly characterizes the score function of denoising posterior distributions. By integrating this insight into the Variational Score Distillation (VSD) framework, we enable scalable learning of generative denoisers capable of approximating samples from the denoising posterior distribution across a wide range of noise levels. The proposed generative denoisers exhibit desirable properties that allow fast generation while preserve the benefit of iterative refinement: (1) fast one-step generation through sampling from pure Gaussian noise at high noise levels; (2) improved sample quality by scaling the test-time compute with multi-step sampling; and (3) zero-shot probabilistic inference for flexible and controllable sampling. We evaluate NCVSD through extensive experiments, including class-conditional image generation and inverse problem solving. By scaling the test-time compute, our method outperforms teacher diffusion models and is on par with consistency models of larger sizes. Additionally, with significantly fewer NFEs than diffusion-based methods, we achieve record-breaking LPIPS on inverse problems.

Training models to efficiently use test-time compute is crucial for improving the reasoning performance of LLMs. While current methods mostly do so via fine-tuning on search traces or running RL against the 0/1 outcome reward, do these approaches efficiently utilize test-time compute? Would these approaches continue to scale as the budget improves? In this paper, we try to answer these questions. We formalize the problem of optimizing test-time compute as a meta reinforcement learning (RL) problem, which provides a principled perspective on spending test-time compute from the lens of exploration and exploitation. It also motivates the use of cumulative regret to measure the efficacy of test-time compute by viewing a long output stream as consisting of several episodes from the model. While current state-of-the-art models do not optimize regret, we show that regret can be minimized by running final 0/1 reward RL regularized by a dense reward bonus, given by the "information gain" from each subsequent block in the output stream. We prescribe an approach for quantifying information gain, which measures the utility of an intermediate segment of tokens towards improving accuracy of the final answer. We instantiate this idea to develop MRT, a new class of finetuning methods for optimizing test-time compute. Fine-tuning with MRT leads to substantial improvements in both performance and token efficiency on the AIME dataset.

Code auditing is the process of reviewing code with the aim of identifying bugs. Large Language Models (LLMs) have demonstrated promising capabilities for this task without requiring compilation, while also supporting user-friendly customization. However, auditing a code repository with LLMs poses significant challenges: limited context windows and hallucinations can degrade the quality of bug reports, and analyzing large-scale repositories incurs substantial time and token costs, hindering efficiency and scalability. This work introduces an LLM-based agent, RepoAudit, designed to perform autonomous repository-level code auditing. Equipped with agent memory, RepoAudit explores the codebase on demand by analyzing data-flow facts along feasible program paths within individual functions. It further incorporates a validator module to mitigate hallucinations by verifying data-flow facts and checking the satisfiability of path conditions associated with potential bugs, thereby reducing false positives. RepoAudit detects 40 true bugs across 15 real-world benchmark projects with a precision of 78.43%, requiring on average only 0.44 hours and $2.54 per project. Also, it detects 185 new bugs in high-profile projects, among which 174 have been confirmed or fixed. We have open-sourced RepoAudit at https://github.com/PurCL/RepoAudit.

Test-Time Adaptation (TTA) enables deep neural networks to adapt to arbitrary distributions during inference. Existing TTA algorithms generally tend to select benign samples that help achieve robust online prediction and stable self-training. Although malicious samples that would undermine the model's optimization should be filtered out, it also leads to a waste of test data. To alleviate this issue, we focus on how to make full use of the malicious test samples for TTA by transforming them into benign ones, and propose a plug-and-play method, PTTA. The core of our solution lies in the purification strategy, which retrieves benign samples having opposite effects on the objective function to perform Mixup with malicious samples, based on a saliency indicator for encoding benign and malicious data. This strategy results in effective utilization of the information in malicious samples and an improvement of the models' online test accuracy. In this way, we can directly apply the purification loss to existing TTA algorithms without the need to carefully adjust the sample selection threshold. Extensive experiments on four types of TTA tasks as well as classification, segmentation, and adversarial defense demonstrate the effectiveness of our method. Code is available at https://github.com/HAIV-Lab/PTTA.

We introduce PokéChamp, a minimax agent powered by Large Language Models (LLMs) for Pokémon battles. Built on a general framework for two-player competitive games, PokéChamp leverages the generalist capabilities of LLMs to enhance minimax tree search. Specifically, LLMs replace three key modules: (1) player action sampling, (2) opponent modeling, and (3) value function estimation, enabling the agent to effectively utilize gameplay history and human knowledge to reduce the search space and address partial observability. Notably, our framework requires no additional LLM training. We evaluate PokéChamp in the popular Gen 9 OU format. When powered by GPT-4o, it achieves a win rate of 76\% against the best existing LLM-based bot and 84\% against the strongest rule-based bot, demonstrating its superior performance. Even with an open-source 8-billion-parameter Llama 3.1 model, PokéChamp consistently outperforms the previous best LLM-based bot, Pokéllmon powered by GPT-4o, with a 64\% win rate. PokéChamp attains a projected Elo of 1300-1500 on the Pokémon Showdown online ladder, placing it among the top 30\%-10\% of human players. In addition, this work compiles the largest real-player Pokémon battle dataset, featuring over 3 million games, including more than 500k high-Elo matches. Based on this dataset, we establish a series of battle benchmarks and puzzles to evaluate specific battling skills. We further provide key updates to the local game engine. This work establishes Pokémon as a benchmark to integrate LLM technologies with game-theoretic algorithms addressing general multi-agent problems. Videos, code, and dataset are available online.

Large language models (LLMs) excel in many reasoning tasks but continue to face significant challenges, such as lack of robustness in reasoning, struggling with cross-task generalization, and inefficiencies in scaling up reasoning capabilities. Current training paradigms, including next-token prediction and reinforcement learning from human feedback, often fall short in adaptability to diverse reasoning tasks. Existing approaches, such as prompt optimization and iterative output refinement, offer performance improvement, but can be inefficient and lack effective generalization. To overcome these limitations, this position paper argues for a transformative shift in how LLMs approach reasoning. Drawing inspiration from cognitive science, particularly meta-reasoning theories such as Dual-Process Theory and Metacognitive Reasoning, we propose a Bayesian meta-reasoning framework for LLMs. Our approach integrates self-awareness, monitoring, evaluation, regulation, and meta-reflection, to enhance LLMs' ability to refine reasoning strategies and generalize across tasks. We revisit existing LLM reasoning methods, identify key challenges, and suggest directions for future research.

Length generalization refers to the ability to extrapolate from short training sequences to long test sequences and is a challenge for current large language models. While prior work has proposed some architecture or data format changes to achieve length generalization, these proposals typically apply to a limited set of tasks. Building on prior scratchpad and Chain-of-Thought (CoT) techniques, we propose *Turing Programs*, a novel CoT strategy that decomposes an algorithmic task into steps mimicking the computation of a Turing Machine. This framework is both universal, as it can accommodate any algorithmic task, and simple, requiring only copying text from the context with small modifications. We show that by using Turing Programs, we obtain robust length generalization on a range of algorithmic tasks: addition, multiplication and in-context SGD. We then demonstrate that transformers achieve length generalization on random Turing Programs, suggesting that length generalization is possible for any algorithmic task. Finally, we theoretically prove that transformers can implement Turing Programs, constructing a simple RASP (Weiss et al.) program that simulates an arbitrary Turing machine.

Recently, newly developed Vision-Language Models (VLMs), such as OpenAI's o1, have emerged, seemingly demonstrating advanced reasoning capabilities across text and image modalities. However, the depth of these advances in language-guided perception and abstract reasoning remains underexplored, and it is unclear whether these models can truly live up to their ambitious promises. To assess the progress and identify shortcomings, we enter the wonderland of Bongard problems, a set of classic visual reasoning puzzles that require human-like abilities of pattern recognition and abstract reasoning. With our extensive evaluation setup, we show that while VLMs occasionally succeed in identifying discriminative concepts and solving some of the problems, they frequently falter. Surprisingly, even elementary concepts that may seem trivial to humans, such as simple spirals, pose significant challenges. Moreover, when explicitly asked to recognize ground truth concepts, they continue to falter, suggesting not only a lack of understanding of these elementary visual concepts but also an inability to generalize to unseen concepts. We compare the results of VLMs to human performance and observe that a significant gap remains between human visual reasoning capabilities and machine cognition.

Large language models can exhibit unexpected behavior in the blink of an eye. In a recent computer use demo, a language model switched from coding to Googling pictures of Yellowstone, and these sudden shifts in behavior have also been observed in reasoning patterns and jailbreaks. This phenomenon is not unique to autoregressive models: in diffusion models, key features of the final output are decided in narrow ``critical windows'' of the generation process. In this work we develop a simple, unifying theory to explain this phenomenon. Using the formalism of stochastic localization for generative models, we show that it emerges generically as the generation process localizes to a sub-population of the distribution it models. While critical windows have been studied at length in diffusion models, existing theory heavily relies on strong distributional assumptions and the particulars of Gaussian diffusion. In contrast to existing work our theory (1) applies to autoregressive and diffusion models; (2) makes very few distributional assumptions; (3) quantitatively improves previous bounds even when specialized to diffusions; and (4) requires basic mathematical tools. Finally, we validate our predictions empirically for LLMs and find that critical windows often coincide with failures in problem solving for various math and reasoning benchmarks.

Tasks that involve complex interactions between objects with unknown dynamics make planning before execution difficult. These tasks require agents to iteratively improve their actions after actively exploring causes and effects in the environment. For these type of tasks, we propose Causal-PIK, a method that leverages Bayesian optimization to reason about causal interactions via a Physics-Informed Kernel to help guide efficient search for the best next action. Experimental results on Virtual Tools and PHYRE physical reasoning benchmarks show that Causal-PIK outperforms state-of-the-art results, requiring fewer actions to reach the goal. We also compare Causal-PIK to human studies, including results from a new user study we conducted on the PHYRE benchmark. We find that Causal-PIK remains competitive on tasks that are very challenging, even for human problem-solvers.

Recent Large Language Models (LLMs) have reported high accuracy on reasoning benchmarks. However, it is still unclear whether the observed results arise from true “reasoning” or from statistical recall of the training set. Inspired by the ladder of causation (Pearl, 2009) and its three levels (associations, interventions and counterfactuals), this paper introduces RE-IMAGINE: a framework to characterize a hierarchy of reasoning ability in LLMs, alongside an automated pipeline to generate problem variations at different levels of the hierarchy. By altering problems in an intermediate symbolic representation, RE-IMAGINE generates arbitrarily many problems that are not solvable using memorization alone. Moreover, the framework is general and can work across reasoning domains, including math, code, and logic. We demonstrate our framework on four widely-used benchmarks to evaluate several families of LLMs, and observe reductions in performance when the models are queried with problem variations. These assessments indicate a degree of reliance on statistical recall for past performance, and open the door to further research targeting skills across the reasoning hierarchy.

Large language models (LLMs) have demonstrated considerable reasoning abilities in various tasks such as mathematics and coding. However, recent studies indicate that even the best models lack true comprehension of their reasoning processes. In this paper, we explore how self-play can enhance the rationality of models in the reasoning process without supervision from humans or superior models. We design a $\textit{\textbf{C}ritic-\textbf{D}iscernment \textbf{G}ame}~(\textbf{CDG})$ in which a prover first provides a solution to a given problem and is subsequently challenged by critiques of its solution. These critiques either aim to assist or mislead the prover. The objective of the prover is to maintain the correct answer when faced with misleading comments, while correcting errors in response to constructive feedback. Our experiments on tasks involving mathematical reasoning, stepwise error detection, self-correction, and long-chain reasoning demonstrate that CDG training can significantly improve the ability of well-aligned LLMs to comprehend their reasoning process.

LLM-as-a-Judge models generate chain-of-thought (CoT) sequences intended to capture the step-by-step reasoning process that underlies the final evaluation of a response. However, due to the lack of human-annotated CoTs for evaluation, the required components and structure of effective reasoning traces remain understudied. Consequently, previous approaches often (1) constrain reasoning traces to hand-designed components, such as a list of criteria, reference answers, or verification questions and (2) structure them such that planning is intertwined with the reasoning for evaluation. In this work, we propose EvalPlanner, a preference optimization algorithm for Thinking-LLM-as-a-Judge that first generates an unconstrained evaluation plan, followed by its execution, and then the final judgment. In a self-training loop, EvalPlanner iteratively optimizes over synthetically constructed evaluation plans and executions, leading to better final verdicts. Our method achieves a new state-of-the-art performance for generative reward models on RewardBench and PPE, despite being trained on fewer amount of, and synthetically generated, preference pairs. Additional experiments on other benchmarks like RM-Bench, JudgeBench, and FollowBenchEval further highlight the utility of both planning and reasoning for building robust LLM-as-a-Judge reasoning models.

Chain-of-thought reasoning and scratchpads have emerged as critical tools for enhancing the computational capabilities of transformers. While theoretical results show that polynomial-length scratchpads can extend transformers' expressivity from $TC^0$ to $PTIME$, their required length remains poorly understood. Empirical evidence even suggests that transformers need scratchpads even for many problems in $TC^0$, such as Parity or Multiplication, challenging optimistic bounds derived from circuit complexity. In this work, we initiate the study of systematic lower bounds for the number of CoT steps across different algorithmic problems, in the hard-attention regime. We study a variety of algorithmic problems, and provide bounds that are tight up to logarithmic factors. Overall, these results contribute to emerging understanding of the power and limitations of chain-of-thought reasoning.

Large language models have demonstrated impressive performance on challenging mathematical reasoning tasks, which has triggered the discussion of whether the performance is achieved by true reasoning capability or memorization. To investigate this question, prior work has constructed mathematical benchmarks when questions undergo simple perturbations -- modifications that still preserve the underlying reasoning patterns of the solutions. However, no work has explored hard perturbations, which fundamentally change the nature of the problem so that the original solution steps do not apply. To bridge the gap, we construct MATH-P-Simple and MATH-P-Hard via simple perturbation and hard perturbation, respectively. Each consists of 279 perturbed math problems derived from level-5 (hardest) problems in the MATH dataset (Hendrycks et al., 2021). We observe significant performance drops on MATH-P-Hard across various models, including o1-mini (-16.49%) and gemini-2.0-flash-thinking (-12.9%). We also raise concerns about a novel form of memorization where models blindly apply learned problem-solving skills without assessing their applicability to modified contexts. This issue is amplified when using original problems for in-context learning. We call for research efforts to address this challenge, which is critical for developing more robust and reliable reasoning models. The project is available at https://math-perturb.github.io/.

Scaling inference compute has become a key driver of advanced reasoning in large language models (LLMs). A proven approach for scaling inference compute is to generate long chains-of-thought (CoTs), enabling models to engage in structured reasoning strategies such as backtracking and error correction. Reinforcement learning (RL) has emerged as a crucial method for developing these capabilities, yet the conditions under which long CoTs emerge remain unclear, and RL training requires careful design choices. In this study, we systematically investigate the underlying *mechanics of long CoT reasoning*—examining the factors that enable models to generate extended reasoning trajectories. Through extensive supervised fine-tuning (SFT) and RL experiments, we identify three key findings: 1) while SFT is not strictly necessary, it significantly simplifies training and improves efficiency; 2) reasoning capabilities tend to emerge with increased training compute but are not guaranteed, making reward shaping essential for stabilizing CoT length growth; and 3) scaling verifiable reward signals is critical for RL, and we find that leveraging noisy, web-extracted solutions with filtering mechanisms shows promising potential, particularly in out-of-distribution (OOD) reasoning tasks such as STEM problem-solving. These insights provide practical guidance for optimizing training strategies to enhance long CoT reasoning in LLMs.