NeurIPS 2025 论文列表

Large reasoning models (LRMs) have demonstrated impressive capabilities in complex problem-solving, yet their internal reasoning mechanisms remain poorly understood. In this paper, we investigate the reasoning trajectories of LRMs from an information-theoretic perspective. By tracking how mutual information (MI) between intermediate representations and the correct answer evolves during LRM reasoning, we observe an interesting MI peaks phenomenon: the MI at specific generative steps exhibits a sudden and significant increase during LRM's reasoning process. We theoretically analyze such phenomenon and show that as MI increases, the probability of model's prediction error decreases. Furthermore, these MI peaks often correspond to tokens expressing reflection or transition, such as "Hmm", "Wait" and "Therefore," which we term as the thinking tokens. We then demonstrate that these thinking tokens are crucial for LRM's reasoning performance, while other tokens has minimal impacts. Building on these analyses, we propose two simple yet effective methods to improve LRM's reasoning performance, by delicately leveraging these thinking tokens. Overall, our work provides novel insights into the reasoning mechanisms of LRMs and offers practical ways to improve their reasoning capabilities. The code is available at \url{https://github.com/ChnQ/MI-Peaks}.

A central question in sensory neuroscience is how much, but also what information neurons transmit about the world. While Shannon’s information theory provides a principled framework to quantify the amount of information neurons encode about all stimuli, it does not reveal which stimuli contribute most, or what stimulus features are encoded. As a concrete example, it is known that neurons in the early visual cortex are 'sensitive' to stimuli in a small region of space (their receptive field). However, it is not clear how such simple intuitions carry to more complex scenarios, e.g. with large, noisy & non-linear population of neurons and high-dimensional stimuli. Several previous measures of neural sensitivity have been proposed. For example, the Fisher information quantifies the sensitivity of neural responses to infinitesimal stimulus perturbations. However, as the Fisher is not a valid decomposition of the mutual information it cannot say how different stimuli contribute to the total encoded information. On the other hand, previous works have proposed stimulus dependent decompositions of mutual information, which define a function $ I(x) $ such that $ I(R; X) = \mathbb{E}[I(x)] $. However, this decomposition is inherently ill-posed: infinitely many functions $I(x)$ satisfy the constraint, with no principled way to select among them. Further, different decompositions behave in qualitatively different ways, making it hard to interpret what are they are telling us. Finally, most proposed decompositions are computationally intractable for the high-dimensional stimuli and non-linear encoding models relevant for neuroscience. To resolve these limitations, we propose a set of axioms that any stimulus specific and feature-specific information decomposition should satisfy in order to serve as a meaningful and interpretable measure of neural sensitivity. These axioms formalize intuitive desiderata: that the information assigned to each stimulus, and stimulus feature, should be non-negative, and additive with respect to repeated measurements. We also require the decomposition to respect a form of locality: changes in how a neuron responds to a stimulus $ x $ should not affect the information attributed to a distant stimulus $ x' $. Finally, the attribution must be insensitive to irrelevant features, which do not contribute to the total information. Together, these constraints ensure that the decomposition is both interpretable and theoretically grounded. We show that existing decompositions violate one or more of these axioms, limiting their interpretability and use as information theoretic measures of neural sensitivity. We then introduce a novel decomposition that satisfies all of our axioms. It generalizes Fisher information by capturing neural sensitivity to both infinitesimal and finite stimulus perturbations. Moreover, it supports further decomposition across individual stimulus features (e.g., image pixels), enabling fine-grained analysis of neural representations. Beyond satisfying our theoretical axioms, our decomposition is computationally tractable for large neural populations and high-dimensional naturalistic stimuli, through the use of diffusion models. We demonstrate the power of our method by quantifying the information encoded by a model of visual neurons about individual images and pixels. Our approach uncovers aspects of the neural code that are not picked up by standard methods, such as the Fisher information, and opens the door to similar analyses in higher-order sensory areas, and artificial neural networks.

The recent rise of Large Reasoning Models (LRMs) has significantly improved multi-step reasoning performance, but often at the cost of generating excessively long reasoning chains. This paper revisits the efficiency of such reasoning processes through an information-theoretic lens, revealing a fundamental trade-off between reasoning length and semantic efficiency. We propose two metrics—InfoBias and InfoGain—to quantify divergence from ideal reasoning paths and stepwise information contribution, respectively. Empirical analyses show that longer reasoning chains tend to exhibit higher information bias and diminishing information gain, especially for incorrect answers. Motivated by these findings, we introduce an entropy-based Adaptive Think strategy that dynamically halts reasoning once confidence is sufficiently high, improving efficiency while maintaining competitive accuracy. Compared to the Vanilla Think approach (default mode), our strategy yields a 1.10% improvement in average accuracy and a 50.80% reduction in token usage on QwQ-32B across six benchmark tasks spanning diverse reasoning types and difficulty levels, demonstrating superior efficiency and reasoning performance. These results underscore the promise of entropy-based methods for enhancing both accuracy and cost-effiiciency in large language model deployment.

We study the problem of computing the exact median by leveraging side information to minimize costly, exact comparisons. We analyze this problem in two key settings: (1) using predictions from unreliable 'weak' oracles, and (2) exploiting known structural information in the form of a partial order. In the classical setting, we introduce a modified LazySelect algorithm that combines weak comparisons with occasional strong comparisons through majority voting. We show that this hybrid strategy has near-linear running time and can achieve high-probability correctness using only sublinear strong comparisons, even when the weak oracle is only slightly better than random guessing. Our theoretical results hold under the persistent comparison model, where resampling will not amplify the probability of correctness. In the partially ordered setting, we generalize the notion of median to directed acyclic graphs (DAGs) and show that the complexity of median selection depends heavily on the DAG's width. We complement our analysis with extensive experiments on synthetic data.

Large Language Models are typically trained on datasets collected from the web, which may inadvertently contain harmful or sensitive personal information. To address growing privacy concerns, unlearning methods have been proposed to remove the influence of specific data from trained models. Of these, exact unlearning---which retrains the model from scratch without the target data---is widely regarded as the gold standard for mitigating privacy risks in deployment. In this paper, we revisit this assumption in a practical deployment setting where both the pre- and post-unlearning logits API are exposed, such as in open-weight scenarios. Targeting this setting, we introduce a novel data extraction attack that leverages signals from the pre-unlearning model to guide the post-unlearning model, uncovering patterns that reflect the removed data distribution. Combining model guidance with a token filtering strategy, our attack significantly improves extraction success rates---doubling performance in some cases---across common benchmarks such as MUSE, TOFU, and WMDP. Furthermore, we demonstrate our attack's effectiveness on a simulated medical diagnosis dataset to highlight real-world privacy risks associated with exact unlearning. In light of our findings, which suggest that unlearning may, in a contradictory way, \textit{increase} the risk of privacy leakage during real-world deployments, we advocate for evaluation of unlearning methods to consider broader threat models that account not only for post-unlearning models but also for adversarial access to prior checkpoints. Code is publicly available at: https://github.com/Nicholas0228/unlearned_data_extraction_llm.

We present an information-theoretic framework for discrete diffusion models that yields principled estimators of log-likelihood using score-matching losses. Inspired by the I-MMSE identity for the Gaussian setup, we derive analogous results for the discrete setting. Specifically, we introduce the Information–Minimum Denoising Score Entropy (I-MDSE) relation, which links mutual information between data and its diffused version to the minimum denoising score entropy (DSE) loss. We extend this theory to masked diffusion and establish the Information–Minimum Denoising Cross-Entropy (I-MDCE) relation, connecting cross-entropy losses to mutual information in discrete masked processes. These results provide a time-integral decomposition of the log-likelihood of the data in terms of optimal score-based losses, showing that commonly used losses such as DSE and DCE are not merely variational bounds but tight and principled estimators of log-likelihood. The I-MDCE decomposition further enables practical extensions, including time-free formula, conditional likelihood estimation in prompt–response tasks, and coupled Monte Carlo estimation of likelihood ratios. Experiments on synthetic and real-world data confirm the accuracy, variance stability, and utility of our estimators. The code is publicly available at https://github.com/Dongjae0324/infodis.

Tabular reasoning requires complex, multi-step information extraction and logical inference, such as aggregation, comparison, or calculation over tabular data. While recent advances have leveraged large language models (LLMs) for reasoning over structured text tables, such high-quality textual representations are often unavailable in real-world settings, where tables typically appear as images. In this paper, we tackle the task of tabular reasoning directly from table images. Our core strategy is to leverage privileged structured information---specifically, the ground-truth structured table data available during training but inaccessible at test time---to enhance multimodal large language models (MLLMs). The key challenges lie in: accurately aligning visual representations with the structured information, particularly mapping the visual evidence to logical steps; and effectively transferring the reasoning skills learned during training to the MLLM for visual inference. To address these, we introduce {\sc Turbo} (TabUlar Reasoning with Bridged infOrmation), a new framework for multimodal tabular reasoning using privileged information. {\sc Turbo} benefits from a structure-aware reasoning trace generator based on DeepSeek-R1, which contributes to high-quality modality-bridged information. On this basis, {\sc Turbo} repeatedly generates and selects advantageous reasoning traces, further enhancing the model's tabular reasoning ability. Experimental results demonstrate that, with limited (9k) data, {\sc Turbo} achieves state-of-the-art performance ($+7.2\%$ vs. previous SOTA) across multiple datasets.

The advent of Machine Learning as a Service (MLaaS) has heightened the trade-off between model explainability and security. In particular, explainability techniques, such as counterfactual explanations, inadvertently increase the risk of model extraction attacks, enabling unauthorized replication of proprietary models. In this paper, we formalize and characterize the risks and inherent complexity of model reconstruction, focusing on the "oracle'' queries required for faithfully inferring the underlying prediction function. We present the first formal analysis of model extraction attacks through the lens of competitive analysis, establishing a foundational framework to evaluate their efficiency. Focusing on models based on additive decision trees (e.g., decision trees, gradient boosting, and random forests), we introduce novel reconstruction algorithms that achieve provably perfect fidelity while demonstrating strong anytime performance. Our framework provides theoretical bounds on the query complexity for extracting tree-based model, offering new insights into the security vulnerabilities of their deployment.

Generative modeling of non-negative, discrete data, such as symbolic music, remains challenging due to two persistent limitations in existing methods. Firstly, many approaches rely on modeling continuous embeddings, which is suboptimal for inherently discrete data distributions. Secondly, most models optimize variational bounds rather than exact data likelihood, resulting in inaccurate likelihood estimates and degraded sampling quality. While recent diffusion-based models have addressed these issues separately, we tackle them jointly. In this work, we introduce the Information-Theoretic Discrete Poisson Diffusion Model (ItDPDM), inspired by photon arrival process, which combines exact likelihood estimation with fully discrete-state modeling. Central to our approach is an information-theoretic Poisson Reconstruction Loss (PRL) that has a provable exact relationship with the true data likelihood. ItDPDM achieves improved likelihood and sampling performance over prior discrete and continuous diffusion models on a variety of synthetic discrete datasets. Furthermore, on real-world datasets such as symbolic music and images, ItDPDM attains superior likelihood estimates and competitive generation quality—demonstrating a proof of concept for distribution-robust discrete generative modeling.

In multimodal representation learning, synergistic interactions between modalities not only provide complementary information but also create unique outcomes through specific interaction patterns that no single modality could achieve alone. Existing methods may struggle to effectively capture the full spectrum of synergistic information, leading to suboptimal performance in tasks where such interactions are critical. This is particularly problematic because synergistic information constitutes the fundamental value proposition of multimodal representation. To address this challenge, we introduce InfMasking, a contrastive synergistic information extraction method designed to enhance synergistic information through an Infinite Masking strategy. InfMasking stochastically occludes most features from each modality during fusion, preserving only partial information to create representations with varied synergistic patterns. Unmasked fused representations are then aligned with masked ones through mutual information maximization to encode comprehensive synergistic information. This infinite masking strategy enables capturing richer interactions by exposing the model to diverse partial modality combinations during training. As computing mutual information estimates with infinite masking is computationally prohibitive, we derive an InfMasking loss to approximate this calculation. Through controlled experiments, we demonstrate that InfMasking effectively enhances synergistic information between modalities. In evaluations on large-scale real-world datasets, InfMasking achieves state-of-the-art performance across seven benchmarks. Code is released at https://github.com/brightest66/InfMasking.

Accurately extracting protocol-state machines (PSMs) from the long, densely written Request-for-Comments (RFC) standards that govern Internet‐scale communication remains a bottleneck for automated security analysis and protocol testing. In this paper, we introduce RFC2PSM, the first large-scale dataset that pairs 1,580 pages of cleaned RFC text with 108 manually validated states and 297 transitions covering 14 widely deployed protocols spanning the data-link, transport, session, and application layers. Built on this corpus, we propose PsmBench, a benchmark that (i) feeds chunked RFC to an LLM, (ii) prompts the model to emit a machine-readable PSM, and (iii) scores the output with structure-aware, semantic fuzzy-matching metrics that reward partially correct graphs. A comprehensive baseline study of nine state-of-the-art open and commercial LLMs reveals a persistent state–transition gap: models identify many individual states (up to $0.82$ F1) but struggle to assemble coherent transition graphs ($\leq 0.38$ F1), highlighting challenges in long-context reasoning, alias resolution, and action/event disambiguation. We release the dataset, evaluation code, and all model outputs as open-sourced, providing a fully reproducible starting point for future work on reasoning over technical prose and generating executable graph structures. RFC2PSM and PsmBench aim to catalyze cross-disciplinary progress toward LLMs that can interpret and verify the protocols that keep the Internet safe.

Recent advancements in language models (LMs) have marked a shift toward the growing importance of post-training. Yet, post-training approaches such as supervised fine-tuning (SFT) do not guarantee the effective use of knowledge acquired during pretraining. We therefore introduce infosteer, a lightweight method that encourages parametric information utilization in LMs during post-training. Specifically, Infosteer treats the feed-forward network (FFN) layer as associate key-value memory and promotes the use of stored memory vectors via forward-pass interventions or regularization during backpropagation. This simple guidance during post-training phase yields consistent performance improvements across diverse model families--including Qwen, Gemma and Llama---spanning 15 downstream tasks in both in-distribution (ID) and out-of-distribution (OOD) evaluations. Beyond performance gains, we also find that steered LMs can adaptively allocate information by placing more emphasis on generating semantically meaningful tokens, while using fewer resources on simple transition ones (e.g., ',' or 'and'). Our work underscores that vanilla post-training does not fully exploit the potential gained during pre-training, and that steering LMs in latent representation space offers a promising approach to enhance both performance and interpretability.

Neural Processes (NPs) combine the adaptability of neural networks with the efficiency of meta-learning, offering a powerful framework for modeling stochastic processes. However, existing methods focus on empirical performance while lacking a rigorous theoretical understanding of generalization. To address this, we propose an information-theoretic framework to analyze the generalization bounds of NPs, introducing dynamical stability regularization to minimize sharpness and improve optimization dynamics. Additionally, we show how noise-injected parameter updates complement this regularization. The proposed approach, applicable to a wide range of NP models, is validated through experiments on classic benchmarks, including 1D regression, image completion, Bayesian optimization, and contextual bandits. The results demonstrate tighter generalization bounds and superior predictive performance, establishing a principled foundation for advancing generalizable NP models.

Generative models that maximize model likelihood have gained traction in many practical settings. Among them, perturbation-based approaches underpin many state-of-the-art likelihood estimation models, yet they often face slow convergence and limited theoretical understanding. In this paper, we derive a tighter likelihood bound for noise-driven models to improve both the accuracy and efficiency of maximum likelihood learning. Our key insight extends the classical Kullback–Leibler (KL) divergence–Fisher information relationship to arbitrary noise perturbations, going beyond the Gaussian assumption and enabling structured noise distributions. This formulation allows flexible use of randomized noise distributions that naturally account for sensor artifacts, quantization effects, and data distribution smoothing, while remaining compatible with standard diffusion training. Treating the diffusion process as a Gaussian channel, we further express the mismatched entropy between data and model, showing that the proposed objective upper-bounds the negative log-likelihood (NLL). In experiments, our models achieve competitive NLL on CIFAR-10 and state-of-the-art results on ImageNet across multiple resolutions, all without data augmentation, and the framework extends naturally to discrete data.

Large language models (LLMs) excel at complex tasks thanks to advances in their reasoning abilities. However, existing methods overlook the trade-off between reasoning effectiveness and efficiency, often encouraging unnecessarily long reasoning chains and wasting tokens. To address this, we propose Learning to Think (L2T), an information-theoretic reinforcement fine-tuning framework for LLMs to make the models achieve optimal reasoning with fewer tokens. Specifically, L2T treats each query-response interaction as a hierarchical session of multiple episodes and proposes a universal dense process reward, i.e., quantifies the episode-wise information gain in parameters, requiring no extra annotations or task-specific evaluators. We propose a method to quickly estimate this reward based on PAC-Bayes bounds and the Fisher information matrix. Theoretical analyses show that it significantly reduces computational complexity with high estimation accuracy. By immediately rewarding each episode's contribution and penalizing excessive updates, L2T optimizes the model via reinforcement learning to maximize the use of each episode and achieve effective updates. Empirical results on various reasoning benchmarks and base models demonstrate the advantage of L2T across different tasks, boosting both reasoning effectiveness and efficiency.

Mutual Information (MI) is a fundamental measure of statistical dependence widely used in representation learning. While direct optimization of MI via its definition as a Kullback-Leibler divergence (KLD) is often intractable, many recent methods have instead maximized alternative dependence measures, most notably, the Jensen-Shannon divergence (JSD) between joint and product of marginal distributions via discriminative losses. However, the connection between these surrogate objectives and MI remains poorly understood. In this work, we bridge this gap by deriving a new, tight, and tractable lower bound on KLD as a function of JSD in the general case. By specializing this bound to joint and marginal distributions, we demonstrate that maximizing the JSD-based information increases a guaranteed lower bound on mutual information. Furthermore, we revisit the practical implementation of JSD-based objectives and observe that minimizing the cross-entropy loss of a binary classifier trained to distinguish joint from marginal pairs recovers a known variational lower bound on the JSD. Extensive experiments demonstrate that our lower bound is tight when applied to MI estimation. We compared our lower bound to state-of-the-art neural estimators of variational lower bound across a range of established reference scenarios. Our lower bound estimator consistently provides a stable, low-variance estimate of a tight lower bound on MI. We also demonstrate its practical usefulness in the context of the Information Bottleneck framework. Taken together, our results provide new theoretical justifications and strong empirical evidence for using discriminative learning in MI-based representation learning.

With the rapid development of artificial intelligence technologies, the demand for training data has surged, exacerbating risks of data leakage. Despite increasing incidents and costs associated with such leaks, data leakage prevention (DLP) technologies lag behind evolving evasion techniques that bypass existing sensitive information recognition (SIR) models. Current datasets lack comprehensive coverage of these adversarial transformations, limiting the evaluation of robust SIR systems. To address this gap, we introduce DataSIR, a benchmark dataset specifically designed to evaluate SIR models on sensitive data subjected to diverse format transformations. We curate 26 sensitive data categories based on multiple international regulations, and collect 131,890 original samples correspondingly. Through empirical analysis of real-world evasion tactics, we implement 21 format transformation methods, which are applied to the original samples, expanding the dataset to 1,647,501 samples to simulate adversarial scenarios. We evaluated DataSIR using four traditional NLP models and four large language models (LLMs). For LLMs, we design structured prompts with varying degrees of contextual hints to assess the impact of prior knowledge on recognition accuracy. These evaluations demonstrate that our dataset effectively differentiates the performance of various SIR algorithms. Combined with its rich category and format diversity, the dataset can serve as a benchmark for evaluating related models and help develop future more advanced SIR models. Our dataset and experimental code are publicly available at https://www.kaggle.com/datasets/fanmo1/datasir and https://github.com/Fan-Mo-ZJU/DataSIR.

Consider the task of locating an unknown target point using approximate distance queries: in each round, a reconstructor selects a reference point and receives a noisy version of its distance to the target. This problem arises naturally in various contexts—from localization in GPS and sensor networks to privacy-aware data access—making it relevant from the perspective of both the reconstructor (seeking accurate recovery) and the responder (aiming to limit information disclosure, e.g., for privacy or security reasons). We study this reconstruction game through a learning-theoretic lens, focusing on the rate and limits of the best possible reconstruction error. Our first result provides a tight geometric characterization of the optimal error in terms of the Chebyshev radius, a classical concept from geometry. This characterization applies to all compact metric spaces (in fact, to all totally bounded spaces) and yields explicit formulas for natural subsets of the Euclidean metric. Our second result addresses the asymptotic behavior of reconstruction, distinguishing between pseudo-finite spaces, where the optimal error is attained after finitely many queries, and spaces where the approximation curve exhibits a nontrivial decay. We characterize pseudo-finiteness for convex subsets of Euclidean spaces.

We introduce H-SPLID, a novel algorithm for learning salient feature representations through the explicit decomposition of salient and non-salient features into separate spaces. We show that H-SPLID promotes learning low-dimensional, task-relevant features. We prove that the expected prediction deviation under input perturbations is upper-bounded by the dimension of the salient subspace and the Hilbert-Schmidt Independence Criterion (HSIC) between inputs and representations. This establishes a link between robustness and latent representation compression in terms of the dimensionality and information preserved. Empirical evaluations on image classification tasks show that models trained with H-SPLID primarily rely on salient input components, as indicated by reduced sensitivity to perturbations affecting non-salient features, such as image backgrounds.