ICML 2026 论文列表

Reinforcement learning (RL) has become a widely adopted technique for improving large language models (LLMs) on complex tasks. Despite this progress, existing RL methods still face challenges in training agents with longer-horizon interactions. One major bottleneck is distinguishing the contribution of different actions in long-horizon interaction, leading to high optimization variance. To address this, we introduce a novel policy gradient method, Hindsight Policy Optimization (HPO), that projects both the current policy distribution and the hindsight distribution into an intent space and extracts low-variance learning signals from the Wasserstein distance between them. We theoretically and empirically show that aggregating semantically similar states and actions in the intent space yields a bounded-variance estimator and improves policy performance stably. Our code is available online.

Anchors is a popular local model-agnostic explanation technique whose applicability is limited by its computational inefficiency. To address this limitation, we propose a memorization-based framework that accelerates Anchors while preserving explanation fidelity and understandability. Our approach leverages the iterative nature of Anchors' algorithm which gradually refines an explanation until it is precise enough for a given input by storing and reusing intermediate results obtained during prior explanations. Specifically, we maintain a memory of low-precision, high-coverage rules and introduce a rule transformation framework to adapt them to new inputs: the horizontal transformation adapts a retrieved explanation to the current input by replacing features, and the vertical transformation refines the general explanation until it is precise enough for the input. We evaluate our method across tabular, text, and image datasets, demonstrating that it significantly reduces explanation generation time while maintaining fidelity and understandability, thereby enabling the practical adoption of Anchors in time-sensitive applications.

While current embodied policies exhibit remarkable manipulation skills, their execution remains unsatisfactorily slow as they inherit the tardy pacing of human demonstrations. Existing acceleration methods typically require policy retraining or costly online interactions, limiting their scalability for large-scale foundation models. In this paper, we propose **S**peed**u**p **P**atch (**SuP**), a lightweight, policy-agnostic framework that enables **plug-and-play acceleration** using solely offline data. SuP introduces an external scheduler that adaptively downsamples action chunks provided by embodied policies to eliminate redundancies. Specifically, we formalize the optimization of our scheduler as a Constrained Markov Decision Process (CMDP) aimed at maximizing efficiency without compromising task performance. Since direct success evaluation is infeasible in offline settings, SuP introduces **World Model based state deviation** as a surrogate metric to enforce safety constraints. By leveraging a learned world model as a virtual evaluator to predict counterfactual trajectories, the scheduler can be optimized via offline reinforcement learning. Empirical results on simulation benchmarks (Libero, Bigym) and real-world tasks validate that SuP achieves an overall $1.8\times$ execution speedup for diverse policies while maintaining their original success rates.

Encoder-only language models are frequently used for a variety of language tasks, including classification and retrieval. However, there has been a lack of recent research efforts for encoder models, especially with respect to multilingual models. We introduce mmBERT, an encoder-only language model pretrained on 3T tokens of multilingual text in over 1800 languages. To build mmBERT we introduce several novel elements to massively multilingual encoder training, including phased data curation and scheduled language inclusion. We add over 1700 low-resource languages to the data mix only during the decay phase, showing that it boosts performance dramatically and maximizes the gains from the relatively small amount of training data without excessive repetition. The model uses recent advances in architecture and training schemes to be faster and more multilingual than other models and we release weights, data, and code. We show that mmBERT significantly outperforms the previous generation, on various tasks, for both high and low-resource languages.

Robust signal enhancement at low SNR is fundamentally challenging because noise becomes strongly entangled with the signal and corrupts local time–frequency (TF) evidence. In this regime, fixed resolution short time Fourier transform (STFT) enhancement with purely data driven convolutional biases can become overconfident in unreliable TF regions, causing unstable suppression or residual artifacts. We propose FracKGMF, which couples Fractional Distance Decay Convolution (FracConv) with Knowledge Guided Multi-view Fusion (KGMF) for expressive TF modeling and reliability aware decisions under heavy corruption. FracConv introduces a lightweight fractional distance decay family that reshapes local interactions into long tailed receptive patterns, enabling aggregation of weak but globally consistent cues when per-bin observations are ambiguous. KGMF uses a wiener inspired reliability prior to calibrate multi-view fusion and reduce excessive suppression in uncertain regions. Experiments on speech and EM benchmarks show consistent improvements over state-of-the-art baselines, with particularly large gains under extremely low SNR, including a 33 dB average improvement on EM signals at -20 dB.

We consider max-min and min-max problems with objective functions that are possibly non-smooth, submodular with respect to the minimiser and concave with respect to the maximiser. We investigate the performance of a zeroth-order method applied to this problem. The method is based on the subgradient of the Lovász extension of the objective function with respect to the minimiser and based on Gaussian smoothing to estimate the smoothed function gradient with respect to the maximiser. In expectation sense, we prove the convergence of the algorithm to an $\epsilon$-saddle point in the offline case. Moreover, we show that, in the expectation sense, in the online setting, the algorithm achieves $O(\sqrt{N\bar{P}_N})$ online duality gap, where $N$ is the number of iterations and $\bar{P}_N$ is the path length of the sequence of optimal decisions. The complexity analysis and hyperparameter selection are presented for all the cases. The theoretical results are illustrated via numerical examples.

Standard factuality evaluations of LLMs treat all errors alike, obscuring whether failures arise from missing knowledge (empty shelves) or from limited access to encoded facts (lost keys). We propose a behavioral framework that profiles factual knowledge at the level of facts rather than questions, characterizing each fact by whether it is encoded, and then by how accessible it is: cannot be recalled, can be directly recalled, or can only be recalled with inference-time computation (thinking). To support such profiling, we introduce WikiProfile, a new benchmark constructed via an automated pipeline with a prompted LLM grounded in web search. Across 4 million responses from 13 LLMs, we find that encoding is nearly saturated in frontier models on our benchmark, with GPT-5 and Gemini-3 encoding 95--98\% of facts. However, recall remains a major bottleneck: many errors previously attributed to missing knowledge instead stem from failures to access it. These failures are systematic and disproportionately affect long-tail facts and reverse questions. Finally, we show that thinking improves recall and can recover a substantial fraction of failures, indicating that future gains may rely less on scaling and more on methods that improve how models utilize what they already encode.

Electrocardiogram (ECG) representation learning via ECG-report alignment is often hindered by the inherent structural and statistical divergence between signals and natural language. Existing methods struggle to bridge this gap with simple contrastive objectives, but struggle with distribution dependencies between heterogeneous features. To address this, we propose **SGERA** (**S**tein-**G**uided **E**CG-**R**eport **A**lignment), which leverages the unique properties of Stein kernels to provide a more rigorous geometric alignment in the latent space: **instance-level** alignment via a Stein-RBF kernel enforces pairwise consistency between ECG and report embeddings and **distribution-level** alignment via a Stein-Score kernel captures higher-order interactions for global alignment. Furthermore, we introduce an ECG-Report matching task with a Hard Sample Mining strategy to refine discriminative boundaries. Experiments across three public datasets demonstrate that SGERA significantly outperforms state-of-the-art SSL methods in zero-shot classification, linear probing, and transfer learning, proving the superiority of Stein-guided alignment in handling complex medical modalities. Code is available at supplementary materials.

Aligning AI systems with human values remains a fundamental challenge, but does our inability to create perfectly aligned models preclude obtaining the benefits of alignment? We study a strategic setting where a human user interacts with multiple differently misaligned AI agents, none of which are individually well-aligned. Our key insight is that when the user’s utility lies approximately within the convex hull of the agents’ utilities, a condition that becomes easier to satisfy as model diversity increases, strategic competition can yield outcomes comparable to interacting with a perfectly aligned model. We model this as a multi-leader Stackelberg game, extending Bayesian persuasion to multi-round conversations between differently informed parties, and prove three results: (1) when perfect alignment would allow the user to learn her Bayes-optimal action, she can also do so in all equilibria under the convex hull condition; (2) under weaker assumptions requiring only approximate utility learning, a non-strategic user employing quantal response achieves near-optimal utility in all equilibria; and (3) when the user selects the best single AI after an evaluation period, equilibrium guarantees remain near-optimal without further distributional assumptions. We complement the theory with two forms of empirical evidence: First, we perform simulations of the best-AI selection game using best response dynamics, which show that competition among individually misaligned agents reliably improves user utility when the approximate convex hull assumption is satisfied, but does not always when it fails. Second, we show that synthetically generated AI utility functions (produced via perturbations of the same prompt to evaluate instances on a movie recommendation (MovieLens) and ethical judgement (ETHICS) dataset) quickly produce a convex hull that contains a good approximation of a given utility function even when none of the individual LLM utility functions is well aligned. We show that this phenomenon extends to human and LLM responses on real-world polling data (OpinionQA): a convex hull of LLM opinions can approximate human opinions more accurately than any individual LLM across a wide range of survey questions.

Consumer-level applications require fast optimization of 3D Gaussian Splatting (3DGS) with high-fidelity novel view rendering. However, existing 3DGS acceleration approaches still incur substantial computation on redundant pixels while sacrificing fine details. In this paper, we present TurboGS, an error-guided training framework that accelerates 3DGS by concentrating optimization on perceptually informative pixels. TurboGS is built upon four core components: (1) a tile-wise sparse pixel sampling, which, driven by multi-view reconstruction errors during training, prioritizes challenging regions and skips well-reconstructed ones to avoid redundant gradient computation; (2) a tile-wise structure-aware loss with sparse Normalized Cross-Correlation, which provides sparse yet effective supervision to preserve fine details and stabilize training; (3) an error-driven Gaussian density control strategy, which dynamically allocates model capacity and removes redundant primitives; and (4) a tailored hybrid optimizer that couples Hessian-informed updates with Adam moment damping to stabilize and improve convergence under sparse supervision. Experiments on standard benchmarks demonstrate that TurboGS can deliver on par or superior rendering quality within 100 seconds (up to 10x training speedup over vanilla 3DGS).

Modern language models have historically relied on two dominant design choices: subword tokenization and autoregressive (AR) ordering. These design decisions bake in priors that dictate a model's learning. Recently, two alternative paradigms have challenged this: byte-level modeling, which bypasses static statistically-derived token vocabularies, and masked diffusion modeling (MDM), which conducts parallel, non-sequential generation. Their intersection represents a fully end-to-end modality-agnostic generative prototype; however, removing these structural priors incurs a significant computational cost. In this work, we investigate this cost through a compute-matched scaling study. Our results reveal that the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes. Our findings from controlled permutation experiments suggest that future modality-agnostic designs must incorporate alternative structural biases to maintain viable scaling trajectories in the byte regime.

Existing EEG models are limited by electrode heterogeneity and rigid "channel-first" architectures that treat sensors as independent features. We propose Brain Signal Rendering (BSR), which reinterprets EEG as a physical projection of neural activity and transforms raw signals into structured spatiotemporal tensors (termed Spectrum Videos), enabling the transfer of rich priors from video foundation models. By utilizing VideoMAE for self-supervised pre-training, BSR learns robust, layout-agnostic spatiotemporal representations that preserve neural topology. We further employ subject-level few-shot learning and introduce cross-montage fine-tuning to rigorously evaluate generalization across subjects and electrode configurations. Experiments show that VideoMAE model integrated with the BSR framework significantly outperforms state-of-the-art spectrum based methods, providing a scalable and data-efficient foundation for generalizable EEG modeling. Our code is available at [https://github.com/yanweifu-sii/BSR-VideoMAE](https://github.com/yanweifu-sii/BSR-VideoMAE).

In multi-view clustering (MVC), conventional anchor learning based models implicitly assume a uniform distribution of anchors across clusters, which could lead to inferior representation, especially when clusters vary significantly in size, as larger clusters require more anchors so as to adequately capture their intrinsic structural complexity. To alleviate this, we design a method termed FCFMVC that explicitly encourages proportional anchor allocation. To be specific, we transfer anchor allocation to discrete sample-cluster learning via bipartite graph bridge, and then backpropagate cluster state consisting of size and dispersion degree to guide anchor assignment. This allows the model to integrate cluster cardinality awareness and structural compactness directly into anchor distribution. On the other hand, we regard anchors as pseudo-samples, introduce an anchor-cluster indicator matrix on each view, and directly constrain the number of anchors assigned to each cluster within a tolerance margin. These two paths are further coupled through anchor-sample label alignment, and collaboratively facilitate anchor generation from fine-grained (anchor-level) to coarse-grained (cluster-level) structures. Besides, the entire optimization operation with linear time and space cost makes FCFMVC well-scalable to large-scale tasks. Experiments on datasets with diverse scales confirm the effectiveness of our FCFMVC.

Training with differential privacy (DP) guarantees dataset members that they cannot be identified by users of the released model. However, those data providers, and, in general, the public, lack methods to efficiently verify that models trained on their data satisfy DP guarantees. The amount of compute needed to verify DP guarantees for current algorithms scales with the amount of computation required to train the model. In this paper we design the first DP algorithm with near optimal privacy-utility trade-offs but whose DP guarantees can be verified cheaper than training. We focus on DP stochastic convex optimization (DP-SCO), where optimal privacy-utility trade-offs are known. Here we show we can obtain tight privacy-utility trade-offs by privately minimizing a series of regularized objectives and only using the standard DP composition bound. Crucially, this method can be verified with much less compute than training. This leads to the first known DP-SCO algorithm with near optimal privacy-utility whose DP verification scales better than training cost, significantly reducing verification costs on large datasets.

Graph Mixture-of-Experts (Graph-MoE) offers a way to scale GNNs via adaptive capacity allocation, with the goal of allowing different experts to capture diverse graph patterns. Its effectiveness heavily depends on the coordination between routing decisions and expert specialization. However, through extensive empirical study, we identify two critical phenomena. First, discrimination loss occurs on both the expert and routing sides, where GNN experts become highly homogenized and the router collapses to a small subset of experts, failing to reflect diverse graph semantics. Second, routing uncertainty is prevalent, as existing routers produce uncertain expert assignments for most nodes, and such uncertainty exhibits a strong negative correlation with model performance. To address these issues, we propose C$^2$GMoE, a novel Graph-MoE framework featuring Contrastive routing and Confidence-aware fusion. We introduce a group-wise contrastive routing strategy that provides explicit guidance for routing optimization by aligning node-level routing decisions with semantic clusters while satisfying load-balancing constraints. Moreover, through a theoretical analysis of generalization error, we develop a confidence-aware fusion mechanism that adaptively reweights expert predictions according to their confidence. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our proposed C$^2$GMoE.

Reinforcement learning (RL) has demonstrated potential for enhancing reasoning in large language models (LLMs). However, effective RL training, which requires medium-difficulty training samples, faces two fundamental challenges: Effective Data Scarcity and Dynamic Difficulty Shifts, where medium-difficulty samples are scarce and become trivial as models improve. Existing methods mitigate this scarcity to some extent by generating training samples. However, these approaches suffer from anchor-free generation, ignoring co-evolution, and difficulty mismatch. To address these issues, we propose D²Evo, a Dual Difficulty-aware self-Evolution RL framework. In each iteration, our method mines medium-difficulty anchors based on the current Solver's capability, trains the Questioner to generate diverse questions at appropriate difficulty levels, and jointly optimizes both components to enable progressive reasoning gains. Extensive experiments demonstrate that D²Evo outperforms existing methods on mathematical reasoning benchmarks with fewer than 2K real mathematical samples, and exhibits strong generalization on general reasoning benchmarks.

Diffusion language models (DLMs) offer substantial speed advantages through parallel decoding, but the lack of token dependencies limits generation quality compared to autoregressive (AR) models. Recent progress attempts to bridge the gap via importance sampling, with DLM being the proposal and AR being the target. However, due to the huge gap between their distributions, the sampling requires a large number of particles and is thus expensive to compute. In this paper, we introduce PoE-Bridge, a novel decoding framework that drastically improves generation speed and accuracy by introducing an intermediate distribution to bridge the gap. The distribution is constructed as a Product-of-Experts (PoE) of the DLM proposal and the AR target. With the intermediate distribution, we first use the DLM to draft multiple continuations in parallel, then apply rejection sampling to verify the drafted tokens and move the resulting candidates toward the PoE. We then use importance sampling to further correct the PoE-aligned candidates toward the AR target. We further propose several improved techniques, including mixed-temperature sampling for enhanced diversity and elastic rejection windows for reducing wasted verification. Empirically, PoE-Bridge achieves significantly improved accuracy with $5\times$ speedup over the standard DLM decoding approach, and recovers at least 95% of the target AR model's performance, efficiently advancing most of the quality gap on challenging mathematical reasoning and coding tasks. Our code is available at [https://github.com/juntongshi48/poe-bridge](https://github.com/juntongshi48/poe-bridge).

The scalability of world models is currently bottlenecked by the scarcity of action annotations. While self-supervised latent action learning offers a potential solution, existing single-view paradigms—relying on information bottlenecks or Vector Quantization (VQ)—often conflate superficial 2D pixel displacements with the underlying physical-spatial dynamics of an action. Consequently, these methods remain highly susceptible to view-dependent noise, such as camera shake. We introduce **MuCoLA** (**Mu**lti-view **Co**nsistent **L**atent **A**ction learning), a framework that learns robust, view-invariant action representations by enforcing semantic consistency across synchronized video streams. MuCoLA utilizes a Student-Teacher network with DINO-style self-distillation to align action distributions across viewpoints, effectively filtering high-frequency visual noise while preserving motion semantics. Theoretical analysis reveals that our multi-view objective functions as a spectral filter, isolating agent dynamics from environmental nuisances. Empirically, MuCoLA significantly outperforms baselines in action regression, video reconstruction, and downstream visual control tasks. Furthermore, we demonstrate that MuCoLA exhibits favorable scaling properties with respect to model capacity and data volume, paving the way for large-scale action-free world modeling.

Time-series analysis is critical in real-world applications, yet the explosion of time-series data imposes severe burdens on storage and computational resources. Recently, dataset condensation has emerged as a promising data-centric solution by synthesizing compact yet informative datasets to replace large-scale raw data. However, existing methods are largely vision-centric, failing to capture unique temporal properties of time series, or task-specific, tightly coupling the condensed data to a particular downstream objective. As a result, these approaches suffer from feature mismatch and fail to generalize across diverse time-series tasks. To bridge this gap, we propose UniTSC, the first unified dataset condensation framework for general time-series analysis. UniTSC employs a multi-view hybrid encoder to capture task-invariant representations across temporal, spectral, and topological perspectives. Building upon this representation, we design a tri-space alignment paradigm that jointly aligns optimization trajectories, power spectral densities, and multivariate dependency structures, enabling comprehensive information preservation under extreme compression. Extensive experiments show that UniTSC retains up to 97.9\% of downstream performance using as little as 0.01\% of the original training data, within our experimental settings tied to standard sequence lengths, revealing that a single batch-equivalent budget ($\textless$ 128 samples) is sufficient to capture the essential dynamics of complex time-series data.