NeurIPS 2025 论文列表

Large Language Models (LLMs) have achieved remarkable performance by capturing complex interactions between input features. To identify these interactions, most existing approaches require enumerating all possible combinations of features up to a given order, causing them to scale poorly with the number of inputs $n$. Recently, Kang et al. (2025) proposed SPEX, an information-theoretic approach that uses interaction sparsity to scale to $n \approx 10^3$ features. SPEX greatly improves upon prior methods but requires tens of thousands of model inferences, which can be prohibitive for large models. In this paper, we observe that LLM feature interactions are often *hierarchical*—higher-order interactions are accompanied by their lower-order subsets—which enables more efficient discovery. To exploit this hierarchy, we propose ProxySPEX, an interaction attribution algorithm that first fits gradient boosted trees to masked LLM outputs and then extracts the important interactions. Experiments across four challenging high-dimensional datasets show that ProxySPEX more faithfully reconstructs LLM outputs by 20\% over marginal attribution approaches while using *$10\times$ fewer inferences* than SPEX. By accounting for interactions, ProxySPEX efficiently identifies the most influential features, providing a scalable approximation of their Shapley values. Further, we apply ProxySPEX to two interpretability tasks. *Data attribution*, where we identify interactions among CIFAR-10 training samples that influence test predictions, and *mechanistic interpretability*, where we uncover interactions between attention heads, both within and across layers, on a question-answering task. The ProxySPEX algorithm is available at <https://github.com/mmschlk/shapiq>.

Machine learning models increasingly map biological sequence-fitness landscapes to predict mutational effects. Effective evaluation of these models requires benchmarks curated from empirical data. Despite their impressive scales, existing benchmarks lack topographical information regarding the underlying fitness landscapes, which hampers interpretation and comparison of model performance beyond averaged scores. Here, we introduce GraphFLA, a Python framework that constructs and analyzes fitness landscapes from diverse modalities (DNA, RNA, protein, and beyond.), accommodating datasets up to millions of mutants. GraphFLA calculates 20 biologically relevant features that characterize 4 fundamental aspects of landscape topography. By applying GraphFLA to over 5,300 landscapes from ProteinGym, RNAGym, and CIS-BP, we demonstrate its utility in interpreting and comparing the performance of dozens of fitness prediction models, highlighting factors influencing model accuracy and respective advantages of different models. Additionally, we release 155 combinatorially complete empirical fitness landscapes, encompassing over 2.2 million sequences across various modalities. All the codes and datasets are available at https://github.com/COLA-Laboratory/GraphFLA.

Current Reinforcement Fine-tuning (RFT) paradigms for Large Language Models (LLMs) suffer from sample inefficiency due to the redundant exposure of identical queries under uniform data sampling. While previous work has explored curriculum learning via heuristic difficulty metrics, these strategies exhibit limitations by neglecting the intrinsic learning signals generated by the model itself, thus leading to suboptimal training regimes. In this paper, we identify a model-inherent signal termed *angle concentration* that effectively reflects an LLM's capacity to learn from specific data. We theoretically and empirically demonstrate a correlation between the angular distribution of token hidden state vectors and the resulting gradient, revealing a learning preference for data exhibiting higher angle concentration. Inspired by this finding, we propose GAIN-RL, a Gradient-driven Angle-Informed Navigated RL framework. By leveraging the model's intrinsic angle concentration signal, GAIN-RL dynamically selects training data in each epoch, ensuring consistently impactful gradient updates and thus significantly enhancing overall training efficiency. Empirical evaluations show that GAIN-RL (GRPO) achieves over a 2.5$\times$ acceleration in training efficiency across diverse mathematical and coding tasks and varying model scales. Furthermore, GAIN-RL (GRPO)'s efficient sampling yields data-efficient training, achieving better performance with half the original data compared to vanilla GRPO with full training data.

Inducing-point-based sparse variational Gaussian processes have become the standard workhorse for scaling up GP models. Recent advances show that these methods can be improved by introducing a diagonal scaling matrix to the conditional posterior density given the inducing points. This paper first considers an extension that employs a block-diagonal structure for the scaling matrix, provably tightening the variational lower bound. We then revisit the unifying framework of sparse GPs based on Power Expectation Propagation (PEP) and show that it can leverage and benefit from the new structured approximate posteriors. Through extensive regression experiments, we show that the proposed block-diagonal approximation consistently performs similarly to or better than existing diagonal approximations while maintaining comparable computational costs. Furthermore, the new PEP framework with structured posteriors provides competitive performance across various power hyperparameter settings, offering practitioners flexible alternatives to standard variational approaches.

Joint Alignment (JA) of images aims to align a collection of images into a unified coordinate frame, such that semantically-similar features appear at corresponding spatial locations. Most existing approaches often require long training times, large-capacity models, and extensive hyperparameter tuning. We introduce FastJAM, a rapid, graph-based method that drastically reduces the computational complexity of joint alignment tasks. FastJAM leverages pairwise matches computed by an off-the-shelf image matcher, together with a rapid nonparametric clustering, to construct a graph representing intra- and inter-image keypoint relations. A graph neural network propagates and aggregates these correspondences, efficiently predicting per-image homography parameters via image-level pooling. Utilizing an inverse-compositional loss, that eliminates the need for a regularization term over the predicted transformations (and thus also obviates the hyperparameter tuning associated with such terms), FastJAM performs image JA quickly and effectively. Experimental results on several benchmarks demonstrate that FastJAM achieves results better than existing modern JA methods in terms of alignment quality, while reducing computation time from hours or minutes to mere seconds. Our code is available at our project webpage, https://bgu-cs-vil.github.io/FastJAM/.

What property of the data distribution determines the excess risk of principal component analysis? In this paper, we provide a precise answer to this question. We establish a central limit theorem for the error of the principal subspace estimated by PCA, and derive the asymptotic distribution of its excess risk under the reconstruction loss. We obtain a non-asymptotic upper bound on the excess risk of PCA that recovers, in the large sample limit, our asymptotic characterization. Underlying our contributions is the following result: we prove that the negative block Rayleigh quotient, defined on the Grassmannian, is generalized self-concordant along geodesics emanating from its minimizer of maximum rotation less than $\pi/4$.

In recent years, many studies proposed to generate solutions for Directed Acyclic Graph (DAG) scheduling problem in one shot by combining reinforcement learning and list scheduling heuristic. However, these existing methods suffer from biased estimation of sampling probabilities and inefficient guidance in training, due to redundant comparisons among node priorities and the sparse reward challenge. To address these issues, we analyze of the limitations of these existing methods, and propose a novel one-shot DAG scheduling method with comparability identification and dense reward signal, based on the policy gradient framework. In our method, a comparable antichain identification mechanism is proposed to eliminate the problem of redundant nodewise priority comparison. We also propose a dense reward signal for node level decision-making optimization in training, effectively addressing the sparse reward challenge. The experimental results show that the proposed method can yield superior results of scheduling objectives compared to other learning-based DAG scheduling methods.

Despite its widespread adoption, Adam's advantage over Stochastic Gradient Descent (SGD) lacks a comprehensive theoretical explanation. This paper investigates Adam's sensitivity to rotations of the parameter space. We observe that Adam's performance in training transformers degrades under random rotations of the parameter space, indicating a crucial sensitivity to the choice of basis in practice. This reveals that conventional rotation-invariant assumptions are insufficient to capture Adam's advantages theoretically. To better understand the rotation-dependent properties that benefit Adam, we also identify structured rotations that preserve or even enhance its empirical performance. We then examine the rotation-dependent assumptions in the literature and find that they fall short in explaining Adam's behavior across various rotation types. In contrast, we verify the orthogonality of the update as a promising indicator of Adam’s basis sensitivity, suggesting it may be the key quantity for developing rotation-dependent theoretical frameworks that better explain its empirical success.

Contrary to common belief, we show that gradient ascent-based unconstrained optimization methods frequently fail to perform machine unlearning, a phenomenon we attribute to the inherent statistical dependence between the forget and retain data sets. This dependence, which can manifest itself even as simple correlations, undermines the misconception that these sets can be independently manipulated during unlearning. We provide empirical and theoretical evidence showing these methods often fail precisely due to this overlooked relationship. For random forget sets, this dependence means that degrading forget set metrics (which, for a retrained model, should mirror test set metrics) inevitably harms overall test performance. Going beyond random sets, we consider logistic regression as an instructive example where a critical failure mode emerges: inter-set dependence causes gradient descent-ascent iterations to progressively diverge from the ideal retrained model. Strikingly, these methods can converge to solutions that are not only far from the retrained ideal but are potentially even further from it than the original model itself, rendering the unlearning process actively detrimental. A toy example further illustrates how this dependence can trap models in inferior local minima, inescapable via finetuning. Our findings highlight that the presence of such statistical dependencies, even when manifest only as correlations, can be sufficient for ascent-based unlearning to fail. Our theoretical insights are corroborated by experiments on complex neural networks, demonstrating that these methods do not perform as expected in practice due to this unaddressed statistical interplay.

Recent studies reveal that vision-language models (VLMs) become more susceptible to harmful requests and jailbreak attacks after integrating the vision modality, exhibiting greater vulnerability than their text-only LLM backbones. To uncover the root cause of this phenomenon, we conduct an in-depth analysis and identify a key issue: multimodal inputs introduce an modality-induced activation shift toward a “safer” direction compared to their text-only counterparts, leading VLMs to systematically overestimate the safety of harmful inputs. We refer to this issue as safety perception distortion. To mitigate such distortion, we propose Activation Shift Disentanglement and Calibration (ShiftDC), a training-free method that decomposes and calibrates the modality-induced activation shift to reduce its impact on safety. By isolating and removing the safety-relevant component, ShiftDC restores the inherent safety alignment of the LLM backbone while preserving the vision-language capabilities of VLMs. Experiments demonstrate that ShiftDC significantly enhances safety alignment without impairing model utility.

Prior work using Masked Autoencoders (MAEs) typically relies on random patch masking based on the assumption that images have significant redundancies across different channels, allowing for the reconstruction of masked content using cross-channel correlations. However, this assumption does not hold in Multi-Channel Imaging (MCI), where channels may provide complementary information with minimal feature overlap. Thus, these MAEs primarily learn local structures within individual channels from patch reconstruction, failing to fully leverage cross-channel interactions and limiting their MCI effectiveness. In this paper, we present ChA-MAEViT, an MAE-based method that enhances feature learning across MCI channels via four key strategies: (1) dynamic channel-patch masking, which compels the model to reconstruct missing channels in addition to masked patches, thereby enhancing cross-channel dependencies and improving robustness to varying channel configurations; (2) memory tokens, which serve as long-term memory aids to promote information sharing across channels, addressing the challenges of reconstructing structurally diverse channels; (3) hybrid token fusion module, which merges fine-grained patch tokens with a global class token to capture richer representations; and (4) Channel-Aware Decoder, a lightweight decoder utilizes channel tokens to effectively reconstruct image patches. Experiments on satellite and microscopy datasets, CHAMMI, JUMP-CP, and So2Sat, show that ChA-MAEViT significantly outperforms state-of-the-art MCI-ViTs by 3.0-21.5%, highlighting the importance of cross-channel interactions in MCI. Our code is publicly available at https://github.com/chaudatascience/cha_mae_vit.

Causal inference is essential for developing and evaluating medical interventions, yet real-world medical datasets are often difficult to access due to regulatory barriers. This makes synthetic data a potentially valuable asset that enables these medical analyses, along with the development of new inference methods themselves. Generative models can produce synthetic data that closely approximate real data distributions, yet existing methods do not consider the unique challenges that downstream causal inference tasks, and specifically those focused on treatments, pose. We establish a set of desiderata that synthetic data containing treatments should satisfy to maximise downstream utility: preservation of (i) the covariate distribution, (ii) the treatment assignment mechanism, and (iii) the outcome generation mechanism. Based on these desiderata, we propose a set of evaluation metrics to assess such synthetic data. Finally, we present STEAM: a novel method for generating Synthetic data for Treatment Effect Analysis in Medicine that mimics the data-generating process of data containing treatments and optimises for our desiderata. We empirically demonstrate that STEAM achieves state-of-the-art performance across our metrics as compared to existing generative models, particularly as the complexity of the true data-generating process increases.

Decoding of visual stimuli from noninvasive neuroimaging techniques such as functional magnetic resonance (fMRI) has advanced rapidly in the last years; yet, most high-performing brain decoding models rely on complicated, non-interpretable latent spaces. In this study we present an interpretable brain decoding framework that inserts a semantic bottleneck into BrainDiffuser, a well established, simple and linear decoding pipeline. We firstly produce a $214-\text{dimensional}$ binary interpretable space $\mathcal{L}$ for images, in which each dimension answers to a specific question about the image (e.g., "Is there a person?", "Is it outdoors?"). A first ridge regression maps voxel activity to this semantic space. Because this mapping is linear, its weight matrix can be visualized as maps of voxel importance for each dimension of $\mathcal{L}$, revealing which cortical regions influence mostly each semantic dimension. A second regression then transforms these concept vectors into CLIP embeddings required to produce the final decoded image, conditioning the BrainDiffuser model. We found that voxel-wise weight maps for individual questions are highly consistent with canonical category-selective regions in the visual cortex (face, bodies, places, words), simultaneously revealing that activation distributions, not merely location, bear semantic meaning in the brain. Visual brain decoding performances are only slightly lower compared to the original BrainDiffuser metrics (e.g., the CLIP similarity is decreased by $\leq 4$% for the four subjects), yet offering substantial gains in interpretability and neuroscientific insights. These results show that our interpretable brain decoding pipeline enables voxel-level analysis of semantic representations in the human brain without sacrificing decoding accuracy.

Generative models trained on natural sequences are increasingly used to predict the effects of genetic variation, enabling progress in therapeutic design, disease risk prediction, and synthetic biology. In the zero-shot setting, variant impact is estimated by comparing the likelihoods of sequences, under the assumption that likelihood serves as a proxy for fitness. However, this assumption often breaks down in practice: sequence likelihood reflects not only evolutionary fitness constraints, but also phylogenetic structure and sampling biases, especially as model capacity increases. We introduce Likelihood-Fitness Bridging (LFB), a simple and general strategy that improves variant effect prediction by averaging model scores across sequences subject to similar selective pressures. Assuming an Ornstein-Uhlenbeck model of evolution, LFB can be viewed as a way to marginalize the effects of genetic drift, although its benefits appear to extend more broadly. LFB applies to existing protein and genomic language models without requiring retraining, and incurs only modest computational overhead. Evaluated on large-scale deep mutational scans and clinical benchmarks, LFB consistently improves predictive performance across model families and sizes. Notably, it reverses the performance plateau observed in larger protein language models, making the largest models the most accurate when combined with LFB. These results suggest that accounting for phylogenetic and sampling biases is essential to realizing the full potential of large sequence models in variant effect prediction.

We study the problem of learning in the presence of an adversary that can corrupt an $\eta$ fraction of the training examples with the goal of causing failure on a specific test point. In the realizable setting, prior work established that the optimal error under such instance-targeted poisoning attacks scales as $\Theta(d\eta)$, where $d$ is the VC dimension of the hypothesis class [Hanneke, Karbasi, Mahmoody, Mehalel, and Moran (NeurIPS 2022)]. In this work, we resolve the corresponding question in the agnostic setting. We show that the optimal excess error is $\widetilde\Theta(\sqrt{d\eta})$, answering one of the main open problems left by Hanneke et al. To achieve this rate, it is necessary to use randomized learners: Hanneke et al.\ showed that deterministic learners can be forced to suffer error close to $1$ even under small amounts of poisoning. Perhaps surprisingly, our upper bound remains valid even when the learner’s random bits are fully visible to the adversary. In the other direction, our lower bound is stronger than standard PAC-style bounds: instead of tailoring a hard distribution separately for each sample size, we exhibit a single fixed distribution under which the adversary can enforce an excess error of $\Omega(\sqrt{d\eta})$ infinitely often.

Code embedding models attract increasing attention due to the widespread popularity of retrieval-augmented generation (RAG) in software development. These models are expected to capture the rich semantic relationships inherent to code, which differ significantly from those found in text. However, existing models remain severely limited due to the scarcity of high-quality training data. In this work, we introduce \textbf{CodeR} (\underline{Code} \underline{R}etrieval), a state-of-the-art embedding model for general-purpose code retrieval. The superior performance of CodeR is built upon \textbf{CodeR-Pile}, a large-scale synthetic dataset constructed under the DRU (Diversity, Reliability, Usability) principle via a novel data synthesis pipeline. To optimize training effectiveness, we propose \textbf{Annealing}, a curriculum learning strategy that enables effective knowledge transfer across heterogeneous sources of data. We evaluate CodeR based on 16 diverse code retrieval tasks, where it significantly outperforms existing baselines and exhibits strong out-of-domain generalization performance. We have publicly released our code and the well-trained model to facilitate further research in this critical area\footnote{\url{https://github.com/FlagOpen/FlagEmbedding/tree/master/research/BGE_Coder}}.

Large language models (LLMs) frequently generate hallucinations—content that deviates from factually inaccurate or deviates from provided context—posing challenges for diagnosis. However, diagnosing the causes of hallucination is challenging due to the complex interplay of underlying causes. This paper introduces a framework to systematically understand the sources of hallucination behavior in large language models. Our key insight is that hallucinations arise when more frequent but non-factual associations outweigh faithful ones. Through theoretical and empirical analyses, we demonstrate that decoder-only transformers effectively function as subsequence embedding models, with the fully-connected layers encoding input-output associations. We propose a tracing algorithm that identifies causal subsequences by analyzing hallucination probabilities across randomized input contexts. Experiments show our method outperforms standard attribution techniques in identifying hallucination causes and is supported by evidence from the model’s training corpus. This work provides a unified perspective on hallucinations and a robust framework for their cause and analysis.

We study the fundamental problem of estimating an unknown discrete distribution $p$ over $d$ symbols, given $n$ i.i.d. samples from the distribution. We are interested in minimizing the KL divergence between the true distribution and the algorithm's estimate. We first construct minimax optimal private estimators. Minimax optimality however fails to shed light on an algorithm's performance on individual (non-worst-case) instances $p$ and simple minimax-optimal DP estimators can have poor empirical performance on real distributions. We then study this problem from an instance-optimality viewpoint, where the algorithm's error on $p$ is compared to the minimum achievable estimation error over a small local neighborhood of $p$. Under natural notions of local neighborhood, we propose algorithms that achieve instance-optimality up to constant factors, with and without a differential privacy constraint. Our upper bounds rely on (private) variants of the Good-Turing estimator. Our lower bounds use additive local neighborhoods that more precisely captures the hardness of distribution estimation in KL divergence, compared to ones considered in prior works.

Explaining black-box models for time series data is critical for the wide-scale adoption of deep learning techniques across domains such as healthcare. Recently, explainability methods for deep time series models have seen significant progress by adopting saliency methods that perturb masked segments of time series to uncover their importance towards the prediction of black-box models. Thus far, such methods have been largely restricted to regular time series. Irregular time series, however, sampled at irregular time intervals and potentially with missing values, are the dominant form of time series in various critical domains (e.g., hospital records). In this paper, we conduct the first evaluation of saliency methods for the interpretation of irregular time series models. We first translate techniques for regular time series into the continuous time realm of irregular time series and show under which circumstances such techniques are still applicable. However, existing perturbation techniques neglect the timing and structure of observed data, e.g., informative missingness when data is not missing at random. Thus, we propose Contimask, a simple framework to also apply non-differentiable perturbations, such as simulating that parts of the data had not been observed using NeuroEvolution. Doing so, we successfully detect how structural differences in the data can bias irregular time series models on a real-world sepsis prediction task where 90% of the data is missing. Source code is available on GitHub.