ICML 2026 论文列表

We propose ***Var**iance **D**riven **E**xploration* (VarDE), a principled approach for pure exploration in *highly stochastic environments*, where the exploration process is dominated by stochastic variance. VarDE is built on a fundamental principle: *sampling effort should be allocated to minimize the uncertainty of the final decision*. We formalize the uncertainty of the final decision through a smooth decision function and derive allocation rules that explicitly capture how stochastic noise in individual components affects the reliability of the final output. We apply this methodology to three core problems of pure exploration -- Best Arm Identification (BAI), Monte Carlo Tree Search (MCTS), and Best-Policy Identification (BPI) -- with theoretical guarantees on variance decay and simple regret. Empirically, we demonstrate consistent and significant improvements of VarDE over existing methods, with especially strong gains in highly stochastic environments.

In this work, we show that Latent Flow-Matching (LFM) models are robust to different types of perturbations, including data reduction and model capacity shrinkage. We characterize this stability by these models' tendency to generate similar outputs under identical noise seeds. We provide a perspective relating this phenomenon to flow matching theory, which indicates that this stability is inherent to the FM objective. We further exploit this stability to derive practical algorithms for more efficient training and inference. Concretely, first, we show that by training LFM models on significantly reduced datasets, performance is preserved, and in compute-constrained regimes, the model converges faster while maintaining quality. This yields multiple advantages, including savings in the training time due to faster convergence, and alleviating annotation effort when training conditional models. Second, LFM stability under architectural shrinkage gives rise to a two-model coarse-to-fine approach, one using a light-weight architecture for the first phase of the FM trajectory, and one with higher capacity for the second, thereby reducing the inference cost substantially. To determine which samples are informative, we introduce three sample-scoring criteria and evaluate them under standard metrics for generative models. Our results are thoroughly evaluated on multiple datasets, demonstrating the practical advantage of this stability, including data savings and a more than two-fold inference speedup while generating comparable outputs.

We present MoST (Mixture of Speech and Text), a novel multimodal large language model that seamlessly integrates speech and text processing through our proposed Modality-Aware Mixture of Experts (MAMoE) architecture. While current multimodal models typically process diverse modality representations with identical parameters—disregarding their inherent representational differences, we introduce specialized routing pathways that direct tokens to modality-appropriate experts based on input type. MAMoE simultaneously enhances modality-specific learning and cross-modal understanding through two complementary components: modality-specific expert groups that capture domain-specific patterns and shared experts that facilitate information transfer between modalities. Building on this architecture, we develop an efficient transformation pipeline that adapts the pretrained MoE language model through strategic post-training on ASR and TTS datasets, followed by fine-tuning with a carefully curated speech-text instruction dataset. A key feature of this pipeline is that it relies exclusively on fully accessible, open-source datasets to achieve strong performance and data efficiency. Comprehensive evaluations across ASR, TTS, audio language modeling, and spoken question answering benchmarks show that MoST consistently outperforms existing models of comparable parameter counts. Our ablation studies confirm that the modality-specific routing mechanism and shared experts design significantly contribute to performance gains across all tested domains. To our knowledge, MoST represents the first fully open-source speech-text LLM built on a Mixture of Experts architecture.

We investigate the relation between end-to-end equivariance and layerwise equivariance in deep neural networks. We prove the following: For a network whose end-to-end function is equivariant with respect to group actions on the input and output spaces, there is a parameter choice yielding the same end-to-end function such that its layers are equivariant with respect to some group actions on the latent spaces. Our result assumes that the parameters of the model are identifiable in an appropriate sense. This identifiability property has been established in the literature for a large class of networks, to which our results apply immediately, while it is conjectural for others. The theory we develop is grounded in an abstract formalism, and is therefore architecture-agnostic. Overall, our results provide a mathematical explanation for the emergence of equivariant structures in the weights of neural networks during training -- a phenomenon that is consistently observed in practice.

Diffusion models have demonstrated remarkable generative capabilities and have also emerged as powerful self-supervised representation learners. However, the connection between these two abilities remains less explored. In this work, drawing inspiration from self-supervised learning (SSL), we introduce a framework for jointly evaluating the representation and generation capabilities of diffusion models. Specifically, we decompose features into invariant and residual components and derive the Invariant Contamination Ratio (ICR), a Fisher-based metric that quantifies how residual variation contaminates invariant signal in feature space. We use this framework to analyze both discriminative and generative behavior of diffusion models. On the representation side, we find that invariance peaks at intermediate noise levels, which also yield the best downstream classification performance. On the generative side, we study how training transitions from genuine generalization to memorization in data-limited regimes, and show that ICR serves as a sensitive training-time indicator of early learning: increasing residual energy along Fisher directions marks the onset of memorization, detectable from training features alone without external evaluators or held-out test sets. Overall, our results show that diffusion models can be monitored from a self-supervised perspective through the geometry of their learned representations.

Fine-grained traffic prediction is critically important for mitigating traffic congestion in key urban areas and for providing lane-change guidance in autonomous vehicles and navigation systems. However, task-specific models are not efficient enough, city-scale pre-trained models often overlook fine-grained requirements, and the demand for extensive computational resources hinders practical deployment. To address this issue, we developed a lightweight pre-training framework, MiniTraffic. This framework leverages abundant road-level data to address lane-level data scarcity through a frequency domain stability augmentation module and captures road-lane correlations via contrastive clustering to construct small-scale graph structures, significantly reducing model parameters. Fine-tuning with minimal target data provides a unified and efficient solution for fine-grained traffic prediction. In multi-granularity traffic prediction tasks across six fine-grained datasets, MiniTraffic demonstrated superior performance compared to existing baselines.

Off-policy reinforcement learning is vulnerable to overestimation bias, which is rooted in the total value uncertainty. However, existing methods typically misaddress this by targeting the epistemic component, neglecting the aleatoric component. We identify for the first time that this oversight fails to contain a massive bias surge, termed **Aleatoric Impulse**. Although transient, this impulse fundamentally derails the learning trajectory, permanently locking the agent into suboptimal policies. To counteract this, we propose **A**leatoric **I**mpulse **D**amping **(AID)**, the first mechanism that models total value uncertainty by disentangling the return variance into epistemic and aleatoric components, followed by their adaptive weighted recombination. Leveraging this derived uncertainty, the critic constructs a pessimistic lower confidence bound to surgically suppress the impulse. Complementing this, the actor utilizes a symmetrical upper confidence bound to drive optimistic exploration, ensuring that the necessary pessimism does not compromise exploration efficiency. We integrate this mechanism into the Distributional Soft Actor-Critic algorithm to establish **DSAC-AID**. Extensive experiments on the high-dimensional Gym-MuJoCo and DeepMind Control Suite benchmarks demonstrate that it achieves state-of-the-art results in final performance.

Large-scale models are pretrained on massive web-crawled datasets containing documents of mixed quality, making data filtering essential. A popular method is Classifier-based Quality Filtering (CQF), which trains a binary classifier to distinguish between pretraining data and a small, high-quality set. It assigns each pretraining document a quality score defined as the classifier's score and retains only the top-scoring ones. We provide an in-depth analysis of CQF. We show that while CQF improves downstream task performance, it does not necessarily enhance language modeling on the high-quality set. Importantly, we find that training on CQF-selected data can outperform training directly on the high-quality set, even when the latter is sufficiently large. This finding alone is particularly striking, given the substantial effort and cost recently devoted to augmenting high-quality data. We explain this paradox by the fact that CQF implicitly filters the high-quality dataset as well as the low-quality one. Finally, we introduce an optimization-driven notion of data quality and demonstrate that it can be reliably estimated using small-scale proxy experiments. Altogether, our results both elucidate the mechanisms behind CQF and deepen our understanding of data selection methods widely used in practice.

Diffusion-based Q-learning has emerged as a powerful paradigm for offline reinforcement learning, but its reliance on multi-step denoising makes both training and inference computationally expensive and brittle. Recent efforts to accelerate diffusion Q-learning toward single-step action generation typically introduce auxiliary networks, policy distillation, or multi-phase training, which frequently compromise simplicity, stability, or performance. To address these limitations, we introduce Bootstrapped Flow Q-Learning (BFQ), a novel framework that enables accurate single-step action generation during both training and inference—without auxiliary networks or distillation procedures. BFQ adopts a divide-and-conquer view of the displacement vector along the flow path: it begins by learning short-range displacements that can be accurately estimated from the Flow Matching marginal velocity, and bootstraps these components to directly learn a noise-to-action mapping in a single step. This formulation eliminates multi-step denoising, resulting in a learning procedure that is substantially faster, simpler, and more robust. Extensive D4RL evaluations show that BFQ improves performance while significantly reducing computational cost compared to multi-step diffusion baselines, demonstrating that single-step action generation suffices for high-performance offline Reinforcement Learning.

Pretraining transformers on long sequences, such as entire code repositories or collections of related documents, is bottlenecked by quadratic attention costs. We present Multipole Semantic Attention (MuSe), which accelerates 64k-context pretraining by 36% while matching baseline loss, requiring no architectural changes. MuSe is a training-time approximation that clusters queries and keys separately in representation space. This yields query-specific summaries that substantially outperform spatial blocking at matched sparsity, while also enabling drop-in compatibility with existing pretrained models—we validate on Llama 3.1-8B and 3.2-1B without retraining. We pretrain language models up to 1B parameters at 64k context on code and scientific documents, confirming that MuSe preserves quality and long-context utilization during training.

Recent progress in spatial reasoning with Multimodal Large Language Models (MLLMs) increasingly leverages geometric priors from 3D encoders. However, most existing integration strategies remain passive: geometry is exposed as a global stream and fused in an indiscriminate manner, which often induces semantic-geometry misalignment and redundant signals. We propose GeoThinker, a framework that shifts the paradigm from passive fusion to active perception. Instead of feature mixing, GeoThinker enables the model to selectively retrieve geometric evidence conditioned on its internal reasoning demands. GeoThinker achieves this through Spatial-Grounded Fusion applied at carefully selected VLM layers, where semantic visual priors selectively query and integrate task-relevant geometry via frame-strict cross-attention, further calibrated by Importance Gating that biases per-frame attention toward task-relevant structures. Comprehensive evaluation results show that GeoThinker sets a new state-of-the-art in spatial intelligence, achieving a peak score of 72.6 on the VSI-Bench. Furthermore, GeoThinker demonstrates robust generalization and significantly improved spatial perception across complex downstream scenarios, including embodied referring and autonomous driving. Our results indicate that the ability to actively integrate spatial structures is essential for next-generation spatial intelligence.

Understanding why gradient-based training in deep networks exhibits strong implicit bias remains challenging, in part because tractable singular-value dynamics are typically available only for balanced deep linear models. We propose an alternative route based on two theoretically grounded and empirically testable signatures of deep Jacobians: depth-induced exponential scaling of ordered singular values and strong spectral separation. Adopting a fixed-gates view of piecewise-linear networks, where Jacobians reduce to products of masked linear maps within a single activation region, we prove the existence of Lyapunov exponents governing the top singular values at initialization, give closed-form expressions in a tractable masked model, and quantify finite-depth corrections. We further show that sufficiently strong separation forces singular-vector alignment in matrix products, yielding an approximately shared singular basis for intermediate Jacobians. Together, these results motivate an approximation regime in which singular-value dynamics become effectively decoupled, mirroring classical balanced deep-linear analyses without requiring balancing. Experiments in fixed-gates settings validate the predicted scaling, alignment, and resulting dynamics, supporting a mechanistic account of emergent low-rank Jacobian structure as a driver of implicit bias.

Large language models (LLMs) require substantial compute, and thus energy, at inference time. While quantizing weights and activations is effective at improving efficiency, naive quantization of LLMs can significantly degrade performance due to large magnitude outliers. This paper describes FPTQuant, which introduces three novel, lightweight, and expressive function-preserving transforms (FPTs) to facilitate quantization of transformers: (1) a mergeable pre-RoPE transform for queries and keys, (2) a mergeable transform for values, (3) a cheap, dynamic scaling transform. By leveraging the equivariances and independencies inherent to canonical transformer operation, we designed these FPTs to maintain the model’s function while shaping the intermediate activation distributions to be more quantization friendly. FPTQuant requires no custom kernels and adds virtually no overhead during inference. The FPTs are trained both locally to reduce outliers, and end-to-end such that the outputs of the quantized and full-precision models match. FPTQuant enables static INT4 quantization with minimal overhead and shows SOTA speed-up of up to 3.9x over FP. Empirically, FPTQuant has an excellent accuracy-speed trade-off—it is performing on par or exceeding most prior work and only shows slightly lower accuracy compared to a method that is up to 29% slower.

Personalized federated LoRA fine tuning has become a key approach to addressing data heterogeneity in distributed fine tuning of large language models (LLMs). Existing methods typically assume homogeneous personalization needs across clients, relying on dual LoRA or personalized calibration schemes. However, they fail to account for the heterogeneity of local personalization requirements and the conflicting optimization objectives in dual LoRA, limiting scalability and performance. To address this, we propose Het-CPFLoRA, a customizable heterogeneous federated LoRA fine tuning algorithm inspired by the decoupling properties of LoRA parameters. We employ a single adapter fine tuning scheme to mitigate conflicts between personalized and generalized optimization, decouple LoRA into generalized and personalized subspaces for local customization, and use SVD compression to integrate cross client generalized knowledge. During inference, we introduce an OOD oriented dynamic mechanism to adjust the weighting between personalized and generalized decoupling knowledge, improving performance on user data. Extensive experiments on two public benchmark datasets show that Het-CPFLoRA outperforms state of the art methods in both personalization and generalization across heterogeneous scenarios. The code will be released as an open-source project.

High dynamic range novel view synthesis (HDR-NVS) remains challenged by geometric artifacts and radiometric distortions under multi-exposure conditions, primarily due to existing methods ignoring exposure and over-relying on color cues. Inspired by the integrated processing of color and structure of the human visual system (HVS), we propose Expo-GS, a novel framework that decomposes HDR-NVS into three interpretable components, namely, Irradiance Field Training, Geometry Field Training, and Interactive Joint Training. Central to Expo-GS is the exposure-aware signed distance function (Expo-SDF), which dynamically reweights geometric supervision via localized exposure reliability estimation, suppressing noisy gradients from unstable regions while enhancing structure learning in well-exposed areas. Building on this, we design an interactive optimization strategy that synchronizes Gaussian primitive growth and pruning with evolving Expo-SDF cues, enabling exposure-aware density control and eliminating hallucinated structures near exposure transitions. Experiments show that Expo-GS significantly outperforms prior methods on both synthetic and real-world datasets. It achieves a peak PSNR of 39.06 dB under HDR settings and up to 41.38 dB in the LDR-OE configuration, excelling in preserving high-frequency textures and maintaining structural consistency.

Subsampled natural gradient descent (SNG) has been used to enable high-precision scientific machine learning, but standard analyses based on stochastic preconditioning fail to provide insight into realistic small-sample settings. We overcome this limitation by instead analyzing SNG as a sketch-and-project method. Motivated by this lens, we discard the usual theoretical proxy which decouples gradients and preconditioners using two independent mini-batches, and we replace it with a new proxy based on squared volume sampling. Under this new proxy the expectation of the SNG direction becomes equal to a preconditioned gradient descent step even in the presence of coupling, leading to (i) global convergence guarantees when using a single mini-batch of any size, and (ii) an explicit characterization of the convergence rate in terms of quantities related to the sketch-and-project structure. These findings in turn yield new insights into small-sample settings, for example by suggesting that the advantage of SNG over SGD is that it can more effectively exploit spectral decay in the model Jacobian. We also extend these ideas to explain a popular structured momentum scheme for SNG, known as SPRING, by showing that it arises naturally from accelerated sketch-and-project methods.

A surge of recent advancements has consistently highlighted the superiority of multimodal learning over unimodal approaches across a variety of tasks. However, the theoretical foundations elucidating this advantage remain underexplored: existing theoretical analyses are often constrained by tight assumptions, and lack empirical validation. In this paper, we link this gap by proposing a novel theoretical framework grounded in \textit{convolutional smoothing}, offering a new perspective on how multimodal learning contributes to a smoother loss landscape compared to unimodal learning. Building upon this theoretical foundation, we introduce a simple yet effective distributional training approach based on stochastic modality pairing instead of fixed pairing; thus, further promoting flatter landscape via convolutional smoothing. Our empirical results across various multimodal datasets demonstrate that multimodal models not only achieve better performance but also exhibit smoother loss landscape, which represent better robustness and generalization.

We study how syntactic and semantic information is encoded in inner layer representations of Large Language Models (LLMs), focusing on the very large DeepSeek-V3. We find that, by averaging hidden-representation vectors of sentences sharing syntactic structure or meaning, we obtain vectors that capture a significant proportion of the syntactic and semantic information contained in the representations. In particular, subtracting these syntactic and semantic ``centroids'' from sentence vectors strongly affects their similarity with syntactically and semantically matched sentences, respectively, suggesting that syntax and semantics are, at least partially, linearly encoded. We also find that the cross-layer encoding profiles of syntax and semantics are different, and that the two signals can to some extent be decoupled, suggesting differential encoding of these two types of linguistic information in LLM representations.

Key–value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear–Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.