NeurIPS 2025 论文列表

This paper studies reinforcement learning from human feedback (RLHF) for aligning large language models with human preferences. While RLHF has demonstrated promising results, many algorithms are highly sensitive to misspecifications in the underlying preference model (e.g., the Bradley-Terry model), the reference policy, or the reward function, resulting in undesirable fine-tuning. To address model misspecification, we propose a doubly robust preference optimization algorithm that remains consistent when either the preference model or the reference policy is correctly specified (without requiring both). Our proposal demonstrates superior and more robust performance than state-of-the-art algorithms, both in theory and in practice. The code is available at https://github.com/DRPO4LLM/DRPO4LLM

Autoregressive (AR) language models generate text one token at a time, which limits their inference speed. Diffusion-based language models offer a promising alternative, as they can decode multiple tokens in parallel. However, we identify a key bottleneck in current diffusion LMs: the \textbf{long decoding-window problem}, where tokens generated far from the input context often become irrelevant or repetitive. Previous solutions like semi-autoregressive address this issue by splitting windows into blocks (sacrificing bidirectionality), but we find that this also leads to \textbf{time-interval expansion problem}, sacrificing the speed. Therefore, semi-AR eliminates the main advantages of diffusion models. To overcome this, we propose Convolutional decoding (\textit{Conv}), a normalization-based method that narrows the decoding window without hard segmentation, leading to better fluency and flexibility. Additionally, we introduce Rejecting Rule-based Fine-Tuning (R2FT), a post-hoc training scheme that better aligns tokens at positions far from context. Our methods achieve state-of-the-art results on open-ended generation benchmarks (e.g., AlpacaEval) among diffusion LM baselines, with significantly lower step size than previous works, demonstrating both speed and quality improvements. The code is available online (\url{https://github.com/ybseo-ac/Conv}).

Transformers have theoretical limitations in modeling certain sequence-to-sequence tasks, yet it remains largely unclear if these limitations play a role in large-scale pretrained LLMs, or whether LLMs might effectively overcome these constraints in practice due to the scale of both the models themselves and their pretraining data. We explore how these architectural constraints manifest after pretraining by studying a family of *retrieval* and *copying* tasks inspired by Liu et al. [2024a]. We use a recently proposed framework for studying length generalization [Huang et al., 2025] to provide guarantees for each of our settings. Empirically, we observe an *induction-versus-anti-induction asymmetry*, where pretrained models are better at retrieving tokens to the right (induction) rather than the left (anti-induction) of a query token. This asymmetry disappears upon targeted fine-tuning if length-generalization is guaranteed by theory. Mechanistic analysis reveals that this asymmetry is connected to the differences in the strength of induction versus anti-induction circuits within pretrained transformers. We validate our findings through practical experiments on real-world tasks demonstrating reliability risks. Our results highlight that pretraining selectively enhances certain transformer capabilities, but does not overcome fundamental length-generalization limits.

State-of-the-art neural network verifiers demonstrate that applying the branch-and-bound (BaB) procedure with fast bounding techniques plays a key role in tackling many challenging verification properties. In this work, we introduce the \emph{linear constraint-driven clipping} framework, a class of scalable and efficient methods to enhance bound propagation verifiers. Under this framework, we develop two novel algorithms that efficiently utilize constraints to 1) reduce portions of the input space that are either verified or irrelevant to a subdomain in the context of branch-and-bound, and 2) directly improve intermediate bounds throughout the network. The process novelly uses linear constraints that are readily available during verification in a highly scalable manner compared to using off-the-shelf linear programming (LP) solvers. This reduction tightens bounds globally and can significantly reduce the number of subproblems handled during BaB. We show our clipping procedures can intuitively and efficiently be incorporated into BaB-based verifiers such as $\alpha, \beta$-CROWN, and is amenable to BaB procedures that split upon the input or activation space. We demonstrate the effectiveness of our procedure on a broad range of benchmarks where, in some instances, we witness a 96\% reduction in the number of subproblems during branch-and-bound, and also achieve state-of-the-art verified accuracy across multiple benchmarks.

Palpation, the use of touch in medical examination, is almost exclusively performed by humans. We investigate a proof of concept for an artificial palpation method based on self-supervised learning. Our key idea is that an encoder-decoder framework can learn a **representation** from a sequence of tactile measurements that contains all the relevant information about the palpated object. We conjecture that such a representation can be used for downstream tasks such as tactile imaging and change detection. With enough training data, it should capture intricate patterns in the tactile measurements that go beyond a simple map of forces -- the current state of the art. To validate our approach, we both develop a simulation environment and collect a real-world dataset of soft objects and corresponding ground truth images obtained by magnetic resonance imaging (MRI). We collect palpation sequences using a robot equipped with a tactile sensor, and train a model that predicts sensory readings at different positions on the object. We investigate the representation learned in this process, and demonstrate its use in imaging and change detection.

Deepfake detection remains a pressing challenge, particularly in real-world settings where smartphone-captured media from digital screens often introduces Moiré artifacts that can distort detection outcomes. This study systematically evaluates state-of-the-art (SOTA) deepfake detectors on Moiré-affected videos—an issue that has received little attention. We collected a dataset of 12,832 videos, spanning 35.64 hours, from Celeb-DF, DFD, DFDC, UADFV, and FF++ datasets, capturing footage under diverse real-world conditions, including varying screens, smartphones, lighting setups, and camera angles. To further examine the influence of Moiré patterns on deepfake detection, we conducted additional experiments using our DeepMoiréFake, referred to as (DMF) dataset, and two synthetic Moiré generation techniques. Across 15 top-performing detectors, our results show that Moiré artifacts degrade performance by as much as 25.4\%, while synthetically generated Moiré patterns lead to a 21.4\% drop in accuracy. Surprisingly, demoiréing methods, intended as a mitigation approach, instead worsened the problem, reducing accuracy by up to 16\%. These findings underscore the urgent need for detection models that can robustly handle Moiré distortions alongside other real-world challenges, such as compression, sharpening, and blurring. By introducing the DMF dataset, we aim to drive future research toward closing the gap between controlled experiments and practical deepfake detection.

Large reasoning models (LRMs) are proficient at generating explicit, step-by-step reasoning sequences before producing final answers. However, such detailed reasoning can introduce substantial computational overhead and latency, particularly for simple problems. To address this over-thinking problem, we explore how to equip LRMs with adaptive thinking capabilities—enabling them to dynamically decide whether or not to engage in explicit reasoning based on problem complexity. Building on R1-style distilled models, we observe that inserting a simple ellipsis ("...") into the prompt can stochastically trigger either a thinking or no-thinking mode, revealing a latent controllability in the reasoning behavior. Leveraging this property, we propose AutoThink, a multi-stage reinforcement learning (RL) framework that progressively optimizes reasoning policies via stage-wise reward shaping. AutoThink learns to invoke explicit reasoning only when necessary, while defaulting to succinct responses for simpler tasks. Experiments on five mainstream mathematical benchmarks demonstrate that AutoThink achieves favorable accuracy–efficiency trade-offs compared to recent prompting and RL-based pruning methods. It can be seamlessly integrated into any R1-style model, including both distilled and further fine-tuned variants. Notably, AutoThink improves relative accuracy by 6.4\% while reducing token usage by 52\% on DeepSeek-R1-Distill-Qwen-1.5B, establishing a scalable and adaptive reasoning paradigm for LRMs. Project Page: https://github.com/ScienceOne-AI/AutoThink.

Condensing large datasets into smaller synthetic counterparts has demonstrated its promise for image classification. However, previous research has overlooked a crucial concern in image recognition: ensuring that models trained on condensed datasets are unbiased towards protected attributes (PA), such as gender and race. Our investigation reveals that dataset distillation fails to alleviate the unfairness towards minority groups within original datasets. Moreover, this bias typically worsens in the condensed datasets due to their smaller size. To bridge the research gap, we propose a novel fair dataset distillation (FDD) framework, namely FairDD, which can be seamlessly applied to diverse matching-based DD approaches (DDs), requiring no modifications to their original architectures. The key innovation of FairDD lies in synchronously matching synthetic datasets to PA-wise groups of original datasets, rather than indiscriminate alignment to the whole distributions in vanilla DDs, dominated by majority groups. This synchronized matching allows synthetic datasets to avoid collapsing into majority groups and bootstrap their balanced generation to all PA groups. Consequently, FairDD could effectively regularize vanilla DDs to favor biased generation toward minority groups while maintaining the accuracy of target attributes. Theoretical analyses and extensive experimental evaluations demonstrate that FairDD significantly improves fairness compared to vanilla DDs, with a promising trade-off between fairness and accuracy. Its consistent superiority across diverse DDs, spanning Distribution and Gradient Matching, establishes it as a versatile FDD approach.

The rapid development of multilingual large language models (LLMs) highlights the need for high-quality, diverse, and well-curated multilingual datasets. In this paper, we introduce DCAD-2000 (Data Cleaning as Anomaly Detection), a large-scale multilingual corpus constructed from newly extracted Common Crawl data and existing multilingual sources. DCAD-2000 covers 2,282 languages, 46.72TB of text, and 8.63 billion documents, spanning 155 high- and medium-resource languages and 159 writing scripts. To overcome the limitations of existing data cleaning approaches, which rely on manually designed heuristic thresholds, we reframe data cleaning as an anomaly detection problem. This dynamic filtering paradigm substantially improves data quality by automatically identifying and removing noisy or anomalous content. By fine-tuning LLMs on DCAD-2000, we demonstrate notable improvements in data quality, robustness of the cleaning pipeline, and downstream performance, particularly for low-resource languages across multiple multilingual benchmarks.

Label prototype learning has emerged as an effective paradigm in Partial Multi-Label Learning (PML), providing a distinctive framework for modeling structured representations of label semantics while naturally filtering noise through prototype-based label confidence estimation. However, existing prototype-based methods face a critical limitation: class prototypes are the biased estimates due to noisy candidate labels, particularly when positive samples are scarce. To this end, we first propose a mutually class prototype alignment strategy bypassing noise interference by introducing two different transformation matrices, which makes the class prototypes learned by the fuzzy clustering and candidate label set mutually alignment for correcting themselves. Such alignment is also passed on to the fuzzy memberships label in turn. In addition, to eliminate noise interference in the candidate label set during the classifier learning, we use the learned permutation matrix to transform the fuzzy memberships label for learning a label reliability indicator matrix accompanied by the candidate label set. This makes the label reliability indicator matrix absolutely prevent the occurrence of numerical values located in non-label and simultaneously eliminate the introduction of incorrect label as much as possible. The resulting indicator matrix guides a robust multi-label classifier training process, jointly optimizing label confidence and classifier parameters. Extensive experiments demonstrate that our proposed model exhibits significant performance advantages over state-of-the-art PML approaches.

Foundation Vision Encoders have become indispensable across a wide range of dense vision tasks. However, their operation at low spatial feature resolutions necessitates subsequent feature decompression to enable full-resolution processing. To address this limitation, we introduce JAFAR, a lightweight and flexible feature upsampler designed to enhance the spatial resolution of visual features from any Foundation Vision Encoder to any target resolution. JAFAR features an attention-based upsampling module that aligns the spatial representations of high-resolution queries with semantically enriched low-resolution keys via Spatial Feature Transform modulation. Despite the absence of high-resolution feature ground truth; we find that learning at low upsampling ratios and resolutions generalizes surprisingly well to much higher scales. Extensive experiments demonstrate that JAFAR recovers intricate pixel-level details and consistently outperforms existing feature upsampling techniques across a diverse set of dense downstream applications.

Large Language Models (LLMs) are increasingly deployed as autonomous agents in multi-agent systems, and promising coordination has been demonstrated in handling complex tasks under predefined roles and scripted workflows. However, significant challenges remain in open-ended environments, where agents are inherently self-interested and explicit coordination guidelines are absent. In such scenarios, misaligned incentives frequently lead to social dilemmas and inefficient collective outcomes. Inspired by how human societies tackle similar coordination challenges—through temporary collaborations like employment or subcontracting—a cooperative workflow \textbf{Shapley-Coop} is proposed. This workflow enables self-interested Large Language Model (LLM) agents to engage in emergent collaboration by using a fair credit allocation mechanism to ensure each agent’s contributions are appropriately recognized and rewarded. Shapley-Coop introduces structured negotiation protocols and Shapley-inspired reasoning to estimate agents’ marginal contributions, thereby enabling effective task-time coordination and equitable post-task outcome redistribution. This results in effective coordination that fosters collaboration while preserving agent autonomy, through a rational pricing mechanism that encourages cooperative behavior. Evaluated in two multi-agent games and a software engineering simulation, Shapley-Coop consistently enhances LLM agent collaboration and facilitates equitable outcome redistribution, accurately reflecting individual contributions during the task execution process.

Learning safe reinforcement learning (RL) policies from offline multi-task datasets without direct environmental interaction is crucial for efficient and reliable deployment of RL agents. Benefiting from their scalability and strong in-context learning capabilities, recent approaches attempt to utilize Decision Transformer (DT) architectures for offline safe RL, demonstrating promising adaptability across varying safety budgets. However, these methods primarily focus on single-constraint scenarios and struggle with diverse constraint configurations across multiple tasks. Additionally, their reliance on heuristically defined Return-To-Go (RTG) inputs limits flexibility and reduces learning efficiency, particularly in complex multi-task environments. To address these limitations, we propose CoPDT, a novel DT-based framework designed to enhance adaptability to diverse constraints and varying safety budgets. Specifically, CoPDT introduces a constraint prioritized prompt encoder, which leverages sparse binary cost signals to accurately identify constraints, and a constraint prioritized Return-To-Go (CPRTG) token mechanism, which dynamically generates RTGs based on identified constraints and corresponding safety budgets. Extensive experiments on the OSRL benchmark demonstrate that CoPDT achieves superior efficiency and significantly enhanced safety compliance across diverse multi-task scenarios, surpassing state-of-the-art DT-based methods by satisfying safety constraints in more than twice as many tasks.

Tokenization efficiency plays a critical role in the performance and cost of large language models (LLMs), yet most models rely on static tokenizers optimized on general-purpose corpora. These tokenizers’ fixed vocabularies often fail to adapt to domain- or language-specific inputs, leading to longer token sequences and higher computational costs. We introduce zip2zip, a novel method for achieving context-adaptive tokenization in LLMs at inference time. Leveraging an online data compression algorithm (Lempel–Ziv–Welch), zip2zip dynamically expands its active vocabulary at inference time by continuously replacing fragmented token sequences with more compact hypertokens, which it can immediately output during generation. In doing so, the model refines its internal tokenization scheme to match the token distribution of the current context, reducing redundancy and improving representational efficiency. zip2zip consists of three key components: (1) a tokenizer based on Lempel–Ziv–Welch compression that incrementally merges co-occurring tokens into reusable hypertokens on the fly; (2) a dynamic embedding (and unembedding) layer that computes embeddings for newly formed hypertokens at runtime; and (3) a variant of autoregressive language modeling that pretrains the model to handle hypertokenized, compressed text sequences as inputs and outputs. We show that an existing LLM can be uptrained for zip2zip in 10 GPU-hours via parameter-efficient finetuning. The resulting LLM performs test-time adaptation, learning to use hypertokens in unseen contexts and reducing input and output tokens by 15–40%. Code and models are released at https://github.com/epfl-dlab/zip2zip.

Diverse decoding of large language models is crucial for applications requiring multiple semantically distinct responses, yet existing methods primarily achieve lexical rather than semantic diversity. This limitation significantly constrains Best-of-N strategies, group-based reinforcement learning, and data synthesis. While temperature sampling and diverse beam search modify token distributions or apply n-gram penalties, they fail to ensure meaningful semantic differentiation. We introduce Semantic-guided Diverse Decoding (SemDiD), operating directly in embedding space that balances quality with diversity through three complementary mechanisms: orthogonal directional guidance, dynamic inter-group repulsion, and position-debiased probability assessment. SemDiD harmonizes these competing objectives using adaptive gain functions and constraint optimization, ensuring both quality thresholds and maximal semantic differentiation. Experiments show SemDiD consistently outperforms existing methods, improving Best-of-N coverage by 1.4-5.2% across diverse tasks and accelerating RLHF training convergence by 15% while increasing accuracy by up to 2.1%.

In enterprise settings, organizational data is segregated, siloed and carefully protected by elaborate access control frameworks. These access control structures can completely break down if an LLM fine-tuned on the siloed data serves requests, for downstream tasks, from individuals with disparate access privileges. We propose Permissioned LLMs (PermLLM), a new class of LLMs that superimpose the organizational data access control structures on query responses they generate. We formalize abstractions underpinning the means to determine whether access control enforcement happens correctly over LLM query responses. Our formalism introduces the notion of a relevant response that can be used to prove whether a PermLLM mechanism has been implemented correctly. We also introduce a novel metric, called access advantage, to empirically evaluate the efficacy of a PermLLM mechanism. We introduce three novel PermLLM mechanisms that build on Parameter Efficient Fine-Tuning to achieve the desired access control. We furthermore present two instantiations of access advantage–(i) Domain Distinguishability Index (DDI) based on Membership Inference Attacks, and (ii) Utility Gap Index (UGI) based on LLM utility evaluation. We demonstrate the efficacy of our PermLLM mechanisms through extensive experiments on five public datasets (GPQA, RCV1, SimpleQA, WMDP, and PubMedQA), in addition to evaluating the validity of DDI and UGI metrics themselves for quantifying access control in LLMs.

Autoregressive models (ARMs) show strong capabilities in many domains but face challenges with planning and complex reasoning due to their sequential generation. Masked diffusion models (MDMs) address these issues by enabling controllable, any-order, and parallel generation but encounter training difficulties as token prediction complexity varies with unmasked token positions. This work identifies two key characteristics of efficient MDM sampling paths: prioritizing tokens near unmasked ones and generating subsequence earlier in reasoning. We propose the Simple Path Masked Diffusion Model (SPMDM), which partitions sequences into fixed-length, non-overlapping subsequences and applies varying noise scales to learn token-level and cross-subsequence dependencies. Experiments on synthetic data and tasks like Countdown and Sudoku show SPMDM captures structural rules effectively, significantly outperforming existing MDMs and ARMs, with competitive results on broader reasoning benchmarks.

According to Algorithmic Information Theory (AIT), intelligent representations compress data into the shortest possible program while remaining predictive of its content—exhibiting low Kolmogorov Complexity (KC). In contrast, most visual representation learning systems assign fixed-length representations to all inputs, ignoring variations in complexity or familiarity. Recent adaptive tokenization methods address this by allocating variable-length representations but typically require test-time search over multiple hypotheses to identify the most predictive one. Inspired by KC principles, we propose a one-shot adaptive tokenizer, KARL, that predicts the appropriate number of tokens for an image in a single forward pass, halting once its approximate KC is reached. The token count serves as a proxy for the minimum description length. KARL performs comparably to recent adaptive tokenizers while operating in a one-pass manner. Additionally, we present a conceptual study showing a correlation between adaptive tokenization and core ideas from AIT. We demonstrate that adaptive tokenization not only aligns with KC but also reveals empirical signals approximating AIT concepts such as sophistication and logical depth. Finally, we analyze predicted image complexity and interestingness across axes such as structure vs. noise and in-distribution vs. out-of-distribution familiarity, highlighting alignment with human annotations.

Recently, self-play fine-tuning (SPIN) has been proposed to adapt large language models to downstream applications with scarce expert-annotated data, by iteratively generating synthetic responses from the model itself. However, SPIN is designed to optimize the current reward advantages of annotated responses over synthetic responses at hand, which may gradually vanish during iterations, leading to \textit{unstable optimization}. Moreover, the utilization of reference policy induces a \textit{misalignment} issue between the reward formulation for training and the metric for generation. To address these limitations, we propose a novel \textbf{T}riplet-based \textbf{S}elf-\textbf{P}lay f\textbf{I}ne-tu\textbf{N}ing (TSPIN) method that integrates two key designs. First, beyond current advantages, TSPIN additionally incorporates historical advantages between iteratively generated responses and proto-synthetic responses produced by the initial policy. Even if the current advantages diminish, historical advantages remain effective, stabilizing the overall optimization. Second, TSPIN introduces the entropy constraint into the self-play framework, which is theoretically justified to support reference-free fine-tuning, eliminating the training-generation discrepancy. Empirical results on various tasks demonstrate not only the superior performance of TSPIN over SPIN, but also its stable evolution during iterations. Remarkably, compared to supervised fine-tuning, TSPIN achieves comparable or even better performance with only $25\\%$ samples, highlighting its effectiveness when faced with scarce annotated data.