NeurIPS 2025 论文列表

We explore the connection between deep learning and information theory through the paradigm of diffusion models. A diffusion model converts noise into structured data by reinstating, imperfectly, information that is erased when data was diffused to noise. This information is stored in a neural network during training. We quantify this information by introducing a measure called \textit{neural entropy}, which is related to the total entropy produced by diffusion. Neural entropy is a function of not just the data distribution, but also the diffusive process itself. Measurements of neural entropy on a few simple image diffusion models reveal that they are extremely efficient at compressing large ensembles of structured data.

Agents are minimally entities that are influenced by their past observations and act to influence future observations. This latter capacity is captured by empowerment, which has served as a vital framing concept across artificial intelligence and cognitive science. This former capacity, however, is equally foundational: In what ways, and to what extent, can an agent be influenced by what it observes? In this paper, we ground this concept in a universal agent-centric measure that we refer to as plasticity, and reveal a fundamental connection to empowerment. Following a set of desiderata on a suitable definition, we define plasticity using a new information-theoretic quantity we call the generalized directed information. We show that this new quantity strictly generalizes the directed information introduced by Massey (1990) while preserving all of its desirable properties. Under this definition, we find that plasticity is well thought of as the mirror of empowerment: The two concepts are defined using the same measure, with only the direction of influence reversed. Our main result establishes a tension between the plasticity and empowerment of an agent, suggesting that agent design needs to be mindful of both characteristics. We explore the implications of these findings, and suggest that plasticity, empowerment, and their relationship are essential to understanding agency.

We study the problem of training neural networks with quantized parameters. Learning low-precision quantized parameters by enabling computation of gradients via the Straight-Through Estimator (STE) can be challenging. While the STE enables back-propagation, which is a first-order method, recent works have explored the use of zeroth-order (ZO) gradient descent for fine-tuning. We note that the STE provides high-quality biased gradients, and ZO gradients are unbiased but can be expensive. We thus propose First-Order-Guided Zeroth-Order Gradient Descent (FOGZO) that reduces STE bias while reducing computations relative to ZO methods. Empirically, we show FOGZO improves the tradeoff between quality and training time in Quantization-Aware Pre-Training. Specifically, versus STE at the same number of iterations, we show a 1-8% accuracy improvement for DeiT Tiny/Small, 1-2% accuracy improvement on ResNet 18/50, and 1-22 perplexity point improvement for LLaMA models with up to 0.3 billion parameters. For the same loss, FOGZO yields a 796$\times$ reduction in computation versus n-SPSA for a 2-layer MLP on MNIST. Code is available at [https://github.com/1733116199/fogzo](https://github.com/1733116199/fogzo).

Target speaker extraction focuses on isolating a specific speaker's voice from an audio mixture containing multiple speakers. To provide information about the target speaker's identity, prior works have utilized clean audio samples as conditioning inputs. However, such clean audio examples are not always readily available. For instance, obtaining a clean recording of a stranger's voice at a cocktail party without leaving the noisy environment is generally infeasible. Limited prior research has explored extracting the target speaker's characteristics from noisy enrollments, which may contain overlapping speech from interfering speakers. In this work, we explore a novel enrollment strategy that encodes target speaker information from the noisy enrollment by comparing segments where the target speaker is talking (Positive Enrollments) with segments where the target speaker is silent (Negative Enrollments). Experiments show the effectiveness of our model architecture, which achieves over 2.1 dB higher SI-SNRi compared to prior works in extracting the monaural speech from the mixture of two speakers. Additionally, the proposed two-stage training strategy accelerates convergence, reducing the number of optimization steps required to reach 3 dB SNR by 60\%. Overall, our method achieves state-of-the-art performance in the monaural target speaker extraction conditioned on noisy enrollments. Our implementation is available at https://github.com/xu-shitong/TSE-through-Positive-Negative-Enroll .

Time normalization is the task of converting natural language temporal expressions into machine-readable representations. It underpins many downstream applications in information retrieval, question answering, and clinical decision-making. Traditional systems based on the ISO-TimeML schema limit expressivity and struggle with complex constructs such as compositional, event-relative, and multi-span time expressions. In this work, we introduce a novel formulation of time normalization as a code generation task grounded in the SCATE framework, which defines temporal semantics through symbolic and compositional operators. We implement a fully executable SCATE Python library and demonstrate that large language models (LLMs) can generate executable SCATE code. Leveraging this capability, we develop an automatic data augmentation pipeline using LLMs to synthesize large-scale annotated data with code-level validation. Our experiments show that small, locally deployable models trained on this augmented data can achieve strong performance, outperforming even their LLM parents and enabling practical, accurate, and interpretable time normalization.

Limited labeling budget severely impedes data-driven research, such as medical analysis, remote sensing and population census, and active inference is a solution to this problem. Prior works utilizing independent sampling have achieved improvements over uniform sampling, but its insufficient usage of available information undermines its statistical efficiency. In this paper, we propose balanced active inference, a novel algorithm that incorporates balanced constraints based on model uncertainty utilizing the cube method for label selection. Under regularity conditions, we establish its asymptotic properties and also prove that the statistical efficiency of the proposed algorithm is higher than its alternatives. Various numerical experiments, including regression and classification in both synthetic setups and real data analysis, demonstrate that the proposed algorithm outperforms its alternatives while guaranteeing nominal coverage.

Information diffusion prediction, which aims to forecast future infected users during the information spreading process on social platforms, is a challenging and critical task for public opinion analysis. With the development of social platforms, mass communication has become increasingly widespread. However, most existing methods based on GNNs and sequence models mainly focus on structural and temporal patterns in social networks, suffering from spurious diffusion connections and insufficient information for the diffusion analysis. We leverage strong reasoning capability of LLMs and develop a LL**M**-based causal framework for d**i**ffusion inf**l**uence **d**erivation (MILD). Comprehensively integrating four key factors of social diffusion, i.e., connections, active timelines, user profiles, and comments, MILD causally infers authentic diffusion links to construct a diffusion influence graph $G_I$. To validate the quality and reliability of our constructed graph $G_I$, we proposed a newly designed set of evaluation metrics w.r.t. diffusion prediction. We show MILD provides a reliable information diffusion structure that 12% absolutely better than the social network structure and achieves the state-of-the-art performance on diffusion prediction. MILD is expected to contribute to high-quality, more explainable, and more trustworthy public opinion analysis.

Controlling the spatial and semantic structure of diffusion-generated images remains a challenge. Existing methods like ControlNet rely on handcrafted condition maps and retraining, limiting flexibility and generalization. Inversion-based approaches offer stronger alignment but incur high inference cost due to dual-path denoising. We present \textbf{FreeControl}, a training-free framework for semantic structural control in diffusion models. Unlike prior methods that extract attention across multiple timesteps, FreeControl performs \textit{one-step attention extraction} from a single, optimally chosen timestep and reuses it throughout denoising. This enables efficient structural guidance without inversion or retraining. To further improve quality and stability, we introduce \textit{Latent-Condition Decoupling (LCD)}: a principled separation of the timestep condition and the noised latent used in attention extraction. LCD provides finer control over attention quality and eliminates structural artifacts. FreeControl also supports compositional control via reference images assembled from multiple sources, enabling intuitive scene layout design and stronger prompt alignment. FreeControl introduces a new paradigm for test-time control—enabling structurally and semantically aligned, visually coherent generation directly from raw images, with the flexibility for intuitive compositional design and compatibility with modern diffusion models at ~5\% additional cost.

This paper investigates the structure of linear operators introduced in Hernandez et al. [2023] that decode specific relational facts in transformer language models. We extend their single-relation findings to a collection of relations and systematically chart their organization. We show that such collections of relation decoders can be highly compressed by simple order-3 tensor networks without significant loss in decoding accuracy. To explain this surprising redundancy, we develop a cross-evaluation protocol, in which we apply each linear decoder operator to the subjects of every other relation. Our results reveal that these linear maps do not encode distinct relations, but extract recurring, coarse-grained semantic properties (e.g., country of capital city and country of food are both in the country-of-X property). This property-centric structure clarifies both the operators' compressibility and highlights why they generalize only to new relations that are semantically close. Our findings thus interpret linear relational decoding in transformer language models as primarily property-based, rather than relation-specific.

We study the task of noiseless linear regression under Gaussian covariates in the presence of additive oblivious contamination. Specifically, we are given i.i.d.\ samples from a distribution $(x, y)$ on $\mathbb R^d \times \mathbb R$ with $x \sim \mathcal N(0,I_d)$ and $y = x^\top \beta + z$, where $z$ is drawn from an unknown distribution that is independent of $x$. Moreover, $z$ satisfies $\mathbb P[z = 0] = \alpha>0$. The goal is to accurately recover the regressor $\beta$ to small $\ell_2$-error. Ignoring computational considerations, this problem is known to be solvable using $O(d/\alpha)$ samples. On the other hand, the best known polynomial-time algorithms require $\Omega(d/\alpha^2)$ samples. Here we provide formal evidence that the quadratic dependence in $1/\alpha$ is inherent for efficient algorithms. Specifically, we show that any efficient Statistical Query algorithm for this task requires VSTAT complexity at least $\tilde{\Omega}(d^{1/2}/\alpha^2)$.

We study the evolution of information in interactive decision making through the lens of a stochastic multi-armed bandit problem. Focusing on a fundamental example where a unique optimal arm outperforms the rest by a fixed margin, we characterize the optimal success probability and mutual information over time. Our findings reveal distinct growth phases in mutual information---initially linear, transitioning to quadratic, and finally returning to linear---highlighting curious behavioral differences between interactive and non-interactive environments. In particular, we show that optimal success probability and mutual information can be decoupled, where achieving optimal learning does not necessarily require maximizing information gain. These findings shed new light on the intricate interplay between information and learning in interactive decision making.

This paper introduces the Quotient Bayesian Learning Rule, an extension of natural-gradient Bayesian updates to probability models that fall outside the exponential family. Building on the observation that many heavy-tailed and otherwise non-exponential distributions arise as marginals of minimal exponential families, we prove that such marginals inherit a unique Fisher–Rao information geometry via the quotient-manifold construction. Exploiting this geometry, we derive the Quotient Natural Gradient algorithm, which takes steepest-descent steps in the well-structured covering space, thereby guaranteeing parameterization-invariant optimization in the target space. Empirical results on the Student-$t$ distribution confirm that our method converges more rapidly and attains higher-quality solutions than previous variants of the Bayesian Learning Rule. These findings position quotient geometry as a unifying tool for efficient and principled inference across a broad class of latent-variable models.

The study of multimodality has garnered significant interest in fields where analyzing interactions among multiple information sources can enhance predictive modeling, data fusion, and interpretability. Partial information decomposition (PID) has emerged as a useful information-theoretic framework to quantify the degree to which individual modalities independently, redundantly, or synergistically convey information about a target variable. However, existing PID methods depend on optimizing over a joint distribution constrained by estimated pairwise probability distributions, which are costly and inaccurate for continuous and high-dimensional modalities. Our first key insight is that the problem can be solved efficiently when the pairwise distributions are multivariate Gaussians, and we refer to this problem as Gaussian PID (GPID). We propose a new gradient-based algorithm that substantially enhances computational efficiency for GPID based on an alternative formulation of the underlying optimization problem. To generalize the applicability to non-Gaussian data, we learn information-preserving encoders to transform random variables of arbitrary input distributions into pairwise Gaussian random variables. Along the way, we resolved an open problem regarding the optimality of joint Gaussian solutions for GPID. Empirical validation on diverse synthetic examples demonstrates that our proposed method provides more accurate and efficient PID estimates than existing baselines. We further evaluate on a series of large-scale multimodal benchmarks to show its utility in real-world applications of quantifying PID in multimodal datasets and selecting high-performing models.

Large Vision-Language Models (LVLMs) are susceptible to hallucinations, where generated responses seem semantically plausible yet exhibit little or no relevance to the input image. Previous studies reveal that this issue primarily stems from LVLMs' over-reliance on language priors while disregarding the visual information during decoding. To alleviate this issue, we introduce a novel Conditional Pointwise Mutual Information (C-PMI) calibrated decoding strategy, which adaptively strengthens the mutual dependency between generated texts and input images to mitigate hallucinations. Unlike existing methods solely focusing on text token sampling, we propose to jointly model the contributions of visual and textual tokens to C-PMI, formulating hallucination mitigation as a bi-level optimization problem aimed at maximizing mutual information. To solve it, we design a token purification mechanism that dynamically regulates the decoding process by sampling text tokens remaining maximally relevant to the given image, while simultaneously refining image tokens most pertinent to the generated response. Extensive experiments across various benchmarks reveal that the proposed method significantly reduces hallucinations in LVLMs while preserving decoding efficiency.

Label shift adaptation aims to recover target class priors when the labelled source distribution $P$ and the unlabelled target distribution $Q$ share $P(X \mid Y) = Q(X \mid Y)$ but $P(Y) \neq Q(Y)$. Classical black‑box shift estimators invert an empirical confusion matrix of a frozen classifier, producing a brittle point estimate that ignores sampling noise and similarity among classes. We present Graph‑Smoothed Bayesian BBSE (GS‑B$^3$SE), a fully probabilistic alternative that places Laplacian–Gaussian priors on both target log‑priors and confusion‑matrix columns, tying them together on a label‑similarity graph. The resulting posterior is tractable with HMC or a fast block Newton–CG scheme. We prove identifiability, $N^{-1/2}$ contraction, variance bounds that shrink with the graph’s algebraic connectivity, and robustness to Laplacian misspecification. We also reinterpret GS‑B$^3$SE through information geometry, showing that it generalizes existing shift estimators.

In real-world information-seeking scenarios, users have dynamic and diverse needs, requiring RAG systems to demonstrate adaptable resilience. To comprehensively evaluate the resilience of current RAG methods, we introduce HawkBench, a human-labeled, multi-domain benchmark designed to rigorously assess RAG performance across categorized task types. By stratifying tasks based on information-seeking behaviors, HawkBench provides a systematic evaluation of how well RAG systems adapt to diverse user needs. Unlike existing benchmarks, which focus primarily on specific task types (mostly factoid queries) and rely on varying knowledge bases, HawkBench offers: (1) systematic task stratification to cover a broad range of query types, including both factoid and rationale queries, (2) integration of multi-domain corpora across all task types to mitigate corpus bias, and (3) rigorous annotation for high-quality evaluation. HawkBench includes 1,600 high-quality test samples, evenly distributed across domains and task types. Using this benchmark, we evaluate representative RAG methods, analyzing their performance in terms of answer quality and response latency. Our findings highlight the need for dynamic task strategies that integrate decision-making, query interpretation, and global knowledge understanding to improve RAG generalizability. We believe HawkBench serves as a pivotal benchmark for advancing the resilience of RAG methods and their ability to achieve general-purpose information seeking.

In this paper, we leverage stochastic projection and lossy compression to establish new conditional mutual information (CMI) bounds on the generalization error of statistical learning algorithms. It is shown that these bounds are generally tighter than the existing ones. In particular, we prove that for certain problem instances for which existing MI and CMI bounds were recently shown in Attias et al. [2024] and Livni [2023] to become vacuous or fail to describe the right generalization behavior, our bounds yield suitable generalization guarantees of the order of $\mathcal{O}(1/\sqrt{n})$, where $n$ is the size of the training dataset. Furthermore, we use our bounds to investigate the problem of data "memorization" raised in those works, and which asserts that there are learning problem instances for which any learning algorithm that has good prediction there exist distributions under which the algorithm must "memorize'' a big fraction of the training dataset. We show that for every learning algorithm, there exists an auxiliary algorithm that does not memorize and which yields comparable generalization error for any data distribution. In part, this shows that memorization is not necessary for good generalization.

Multimodal federated learning (MMFL) has gained increasing popularity due to its ability to leverage the correlation between various modalities, meanwhile preserving data privacy for different clients. However, recent studies show that correlation between modalities increase the vulnerability of federated learning against Gradient Inversion Attack (GIA). The complicated situation of MMFL privacy preserving can be summarized as follows: 1) different modality transmits different amounts of information, thus requires various protection strength; 2) correlation between modalities should be taken into account. This paper introduces an information theory perspective to analyze the leaked privacy in process of MMFL, and tries to propose a more reasonable protection method \textbf{Sec-MMFL} based on assessing different information leakage possibilities of each modality by conditional mutual information and adjust the corresponding protection strength. Moreover, we use mutual information to reduce the cross-modality information leakage in MMFL. Experiments have proven that our method can bring more balanced and comprehensive protection at an acceptable cost.

Open-vocabulary semantic segmentation assigns every pixel a label drawn from an open-ended, text-defined space. Vision–language models such as CLIP excel at zero-shot recognition, yet their image-level pre-training hinders dense prediction. Current approaches either fine-tune CLIP—at high computational cost—or adopt training-free attention refinements that favor local smoothness while overlooking global semantics. In this paper, we present OPMapper, a lightweight, plug-and-play module that injects both local compactness and global connectivity into attention maps of CLIP. It combines Context-aware Attention Injection, which embeds spatial and semantic correlations, and Semantic Attention Alignment, which iteratively aligns the enriched weights with textual prompts. By jointly modeling token dependencies and leveraging textual guidance, OPMapper enhances visual understanding. OPMapper is highly flexible and can be seamlessly integrated into both training-based and training-free paradigms with minimal computational overhead. Extensive experiments demonstrate its effectiveness, yielding significant improvements across 8 open-vocabulary segmentation benchmarks.