ICML 2026 论文列表

Diffusion Language Models (DLMs) have recently emerged as a strong alternative to autoregressive language models (AR-LMs), due to their comparable accuracy and faster inference speed via parallel decoding. However, standard DLM decoding strategies, which rely on unmasking only high-confidence tokens, encounter an inherent information-theoretic bottleneck that restricts decoding progress and ultimately slows down generation. We demonstrate this through an information-theoretic lower bound that the number of decoding rounds must grow linearly with the sample's total information and inversely with the per-round information budget, establishing a bits-to-rounds principle. Motivated by this theory, we propose Explore-Then-Exploit (ETE), a training-free decoding strategy that maximizes information throughput and decoding efficiency. ETE combines cross-block decoding with targeted exploration of high-uncertainty tokens to reshape the conditional distribution and trigger cascades of confident predictions. Experiments across diverse benchmarks verify our theoretical bounds and demonstrate that ETE consistently reduces the number of decoding rounds compared to confidence-only baselines without compromising generation quality. Furthermore, ETE integrates efficiently with KV caching, translating these algorithmic gains into improved tokens-per-second throughput.

Deep sequence models are said to store atomic facts predominantly in the form of *associative* memory: a brute-force lookup of co-occurring entities. We identify a dramatically different form of storage of atomic facts that we term as *geometric* memory. Here, the model has synthesized embeddings encoding novel *global* relationships between all entities, including ones that do not co-occur in training. Such storage is powerful: for instance, we show how it transforms a hard reasoning task involving an $\ell$-fold composition into an easy-to-learn $1$-step navigation task. From this phenomenon, we extract fundamental aspects of neural embedding geometries that are hard to explain. We argue that the rise of such a geometry, as against a lookup of local associations, cannot be straightforwardly attributed to typical supervisory, architectural, or optimizational pressures. Counterintuitively, a geometry is learned even when it is more complex than the brute-force lookup. Then, by analyzing a connection to Node2Vec, we demonstrate how the geometry stems from a spectral bias that—in contrast to prevailing theories—indeed arises naturally despite the lack of various pressures. This analysis also points out to practitioners a visible headroom to make Transformer memory more strongly geometric. We hope the geometric view of parametric memory encourages revisiting the default intuitions that guide researchers in areas like knowledge acquisition, capacity, discovery, and unlearning.

Multimodal Large Language Models (MLLMs) must resolve conflicts when modalities provide contradictory information, a behavior we term "modality following". We propose a framework that decomposes this behavior into case-specific relative preference uncertainty and stable inherent preference. Across diverse MLLMs and benchmarks, the probability of following a modality consistently decreases as its relative preference uncertainty increases, a trend robust to alternative uncertainty indices. This regularity defines a "balance point'' where modality preferences are evenly matched, offering a capability-disentangled measure of modality bias. Layer-wise probing further shows that ambiguous cases near the balance point trigger middle-to-late-layer "concept oscillations," where top predictions vacillate between modality-supported answers. Finally, we demonstrate the framework's utility for preference steering through Supervised Fine-Tuning (SFT). We find that data efficiency is governed by preference uncertainty: training on easy samples (where one modality dominates) fails to generalize, whereas targeting the identified ``boundary cases" is essential for robust preference alignment and suppressing internal vacillation.

Speculative decoding mitigates the memory-bound nature of LLM decoding by using a lightweight draft model to propose multiple tokens for parallel verification. However, its adoption has been limited by the lack of high-quality draft models and scalable training infrastructure. We introduce SpecForge, an open-source and efficient framework for training speculative decoding models with full support for EAGLE-3. SpecForge incorporates target–draft decoupling, hybrid parallelism, optimized training kernels, and tight integration with production-grade inference engines, enabling up to 9.9x faster EAGLE-3 training for Qwen3-235B-A22B compared to the baseline. We further release SpecBundle, a suite of production-grade EAGLE-3 draft models trained with SpecForge for mainstream open-source LLMs, achieving up to 4.48x end-to-end inference speedup on SGLang and addressing the scarcity of high-quality drafts. Finally, we distill a systematic study of speculative decoding training into practical and actionable recipes to guide real-world adoption.

Transfer across diverse task compositions and unseen behaviors remains a significant challenge for vision-language action (VLA) models. Skills are repeatable and atomic components for various tasks, and similarities shared with different skills provide evidence for transferability across behaviors. However, existing skill-centric methods have two problems. First, skills are often loosely organized, lacking a hierarchy that can capture similarities and differences across skills. Second, they lack a mechanism which has the capacity to express transferable skill attributes in a structured parametric space. To this end, we propose SkillNet, which models skill attributes in a hierarchical way and regulates compositional model structure with transferable skill attributes. SkillNet exploits motion code and VerbNet Framework to explicitly model similarities of skills on mechanical properties and semantic roles, and organizes skills in a hierarchical way. Based on this hierarchy, SkillNet leverages the scalability of the mixture-of-experts (MoE) mechanism and develops skill embeddings as soft constraints to enable compositional generalization via similar expert activations on similar skills. On zero-shot and few-shot transfer experiments in simulators and real-world environments, SkillNet achieves an improvement of performance by 16.0% and 23.9%. Meanwhile, SkillNet achieves state-of-the-art performance on in-domain settings.

Retrosynthesis prediction aims to infer the reactant molecules based on a given product molecule, which is a fundamental task in chemical synthesis. However, existing methods rely on a static pattern-matching paradigm, which limits their ability to perform effective logical decision-making from chemical data, leading to a black-box process. We propose Retro-Expert, an interpretable retrosynthesis framework that performs collaborative reasoning by combining the complementary strengths of Large Language Models and specialized models via pure reinforcement learning. It outputs natural language explanations grounded in chemical logic through three components: (1) specialized models provide chemical knowledge that is distilled into a high-quality chemical decision space, (2) LLM-driven critical reasoning to generate predictions with an interpretable reasoning path, and (3) knowledge-grounded policy optimization refines the interpretable decision policy. Experiments show that Retro-Expert surpasses both LLM-based and specialized models across different metrics, while generating chemically grounded explanations that enhance chemists' trust in practice. The source code for this paper is available at https://github.com/MagixRab-ll/Retro-Expert

Reinforcement Learning (RL) with rubric-based rewards has recently shown remarkable progress in enhancing general reasoning capabilities of Large Language Models (LLMs), yet still suffers from ineffective exploration confined to current policy distribution. In fact, RL optimization can be viewed as steering the policy toward an ideal distribution that maximizes the rewards, while effective exploration should align efforts with desired target. Leveraging this insight, we propose HeRL, a ****H***indsight ***e***xperience guided ***R***einforcement ***L***earning* framework to bootstrap effective exploration by explicitly *telling LLMs the desired behaviors* specified in rewards. Concretely, HeRL treats failed attempts along with their unmet rubrics as hindsight experience, which serves as in-context guidance for the policy to explore desired responses beyond its current distribution. Additionally, we introduce a bonus reward to incentivize responses with greater potential for improvement under such guidance. HeRL facilitates effective learning from desired high-quality samples without repeated trial-and-error from scratch, yielding a more accurate estimation of the expected gradient theoretically. Extensive experiments across various benchmarks demonstrate that HeRL achieves superior performance gains over baselines, and can further benefit from experience guided self-improvement at test time.

Hypergraph contrastive learning has shown strong potential for modeling high-order relations, but most existing methods implicitly assume that nodes within the same hyperedge are semantically homogeneous. In real-world hypergraphs, this assumption often breaks down: heterogeneous node--hyperedge interactions may introduce noisy signals during augmentation and message passing, leading to feature contamination and false-positive contrastive alignment. To address this issue, we propose H$^2$CL, a heterogeneity-aware hypergraph contrastive learning framework for robust representation learning. H$^2$CL first estimates node--hyperedge heterogeneity from input features and uses it to guide a Heterogeneity-Aware View Generator, which selectively masks high-heterogeneity incidences and constructs cleaner contrastive views. It further introduces a Heterogeneity-Aware Hypergraph Encoder that dynamically reweights information propagation in both node-to-hyperedge and hyperedge-to-node aggregation, enabling hyperedges to aggregate more homogeneous signals while suppressing heterogeneous noise. We also provide theoretical analysis showing that the encoder corresponds to a coordinate descent step for minimizing a heterogeneity-weighted Dirichlet energy. Extensive experiments on standard benchmarks and larger-scale hypergraphs demonstrate that H$^2$CL achieves competitive or superior performance compared with recent baselines, remains robust under structural noise, and learns reweighting patterns that effectively reduce heterogeneity during message passing. Our code is available at: https://github.com/sxu-yaokx/HHCL

We study an algorithm-independent, worst-case lower bound for the Gaussian process (GP) bandit problem in the frequentist setting, where the reward function is fixed and has a bounded norm in the known reproducing kernel Hilbert space (RKHS). Specifically, we focus on the squared exponential (SE) kernel, one of the most widely used kernel functions in GP bandits. One of the remaining open questions for this problem is the gap in the *dimension-dependent* logarithmic factors between upper and lower bounds. This paper partially resolves this open question under a hyperspherical input domain. We show that any algorithm suffers $\Omega(\sqrt{T (\ln T)^{d} (\ln \ln T)^{-d}})$ cumulative regret, where $T$ and $d$ represent the total number of steps and the dimension of the hyperspherical domain, respectively. Regarding the simple regret, we show that any algorithm requires $\Omega(\epsilon^{-2}(\ln \frac{1}{\epsilon})^d (\ln \ln \frac{1}{\epsilon})^{-d})$ time steps to find an $\epsilon$-optimal point. We also provide the improved $O((\ln T)^{d+1}(\ln \ln T)^{-d})$ upper bound on the maximum information gain for the SE kernel. Our results guarantee the optimality of the existing best algorithm up to *dimension-independent* logarithmic factors under a hyperspherical input domain.

The technique of combining multiple votes to enhance the quality of a decision is the core of boosting algorithms in machine learning. In particular, boosting provably increases decision quality by combining multiple "weak learners"—hypotheses that are only slightly better than random guessing—into a single "strong learner" that classifies data well. There exist various versions of boosting algorithms, which we improve upon through the introduction of QuantumBoost. Inspired by classical work by Barak, Hardt and Kale, our QuantumBoost algorithm achieves the best known runtime over other boosting methods through two innovations. First, it uses a quantum algorithm to compute approximate Bregman projections faster. Second, it combines this with a lazy projection strategy, a technique from convex optimization where projections are performed infrequently rather than every iteration. To our knowledge, QuantumBoost is the first algorithm, classical or quantum, to successfully adopt a lazy projection strategy in the context of boosting.

Tokenization is used almost universally by modern language models, enabling efficient text representation using multi-byte or multi-character tokens. However, prior work has shown that tokenization can introduce distortion into the model's generations, an issue known as the Prompt Boundary Problem (PBP). For example, users are often advised not to end their prompts with a space because it prevents the model from including the space as part of the next token. While this heuristic is effective in English, the underlying PBP continues to affect code generation and languages such as Chinese, where tokens often do not line up with word and syntactic boundaries. In this work, we present an inference-time method to convert any autoregressive LM with a BPE tokenizer into a character-level or byte-level LM. Our method efficiently solves the PBP and is also able to unify the vocabularies of language models with different tokenizers, allowing one to ensemble LMs with different tokenizers at inference time or transfer the post-training from one model to another using proxy-tuning.

Constraint-based causal discovery relies on conditional independence (CI) tests whose reliability degrades as conditioning sets grow, particularly in hub-dominated graphs. Existing methods constrain adjacency or global structure, but leave conditioning-set dimensionality uncontrolled. In this paper, we propose Ordering-Constrained Markov Blanket discovery (OCMB), a paradigm that treats conditioning-set dimensionality as a first-class constraint. OCMB decouples discovery into two stages: lightweight global ordering estimation providing directional priors, followed by local Markov blanket validation within small, ordering-constrained candidate sets. By enforcing directional constraints before any CI test, OCMB ensures bounded conditioning sets even with hub nodes. We show that OCMB recovers correct parent sets provided a high-recall ordering assumption holds, without requiring the ordering to be globally correct. Experiments demonstrate that OCMB significantly improves precision and robustness over constraint-based and hybrid methods in high-dimensional regimes where conventional CI-based approaches fail.

Score-based diffusion models are a highly effective method for generating samples from a distribution of images. We consider scenarios where the training data comes from a noisy version of the target distribution, and present an efficiently implementable modification of the inference procedure to generate noiseless samples. Our approach is motivated by the manifold hypothesis, according to which meaningful data is concentrated around some low-dimensional manifold of a high-dimensional ambient space. The central idea is that noise manifests as low magnitude variation in off-manifold directions in contrast to the relevant variation of the desired distribution which is mostly confined to on-manifold directions. We introduce the notion of an extended score and show that, in a simplified setting, it can be used to reduce small variations to zero, while leaving large variations mostly unchanged. We describe how its approximation can be computed efficiently from an approximation to the standard score and demonstrate its efficacy on toy problems, synthetic data, and real data.

The Neural Tangent Kernel is a theoretical framework for understanding the training dynamics of neural networks. However, standard NTK and its variants fail to properly depict the finetuning of foundation models, as they neglect the preconditioning effects of adaptive gradients. To bridge this gap, we propose the Optimizer Aware Kernel (OAK), which incorporates the optimizer's influence into standard NTK framework by a preconditioner estimation technique. Furthermore, we conduct an analysis to answer: when and why kernel regime fails in finetuning. We derive explicit error bounds showing that the collapse of kernel regime is primarily due to the cumulative training effects and the task discrepancy between pretraining and finetuning. Theoretically, we justify OAK's preconditioner estimation by bounding its error term. Empirically, experiments on various model architectures show both the effectiveness of the OAK method and validity of our arguments on kernel regime collapse.

World-model-based imagine-then-act becomes a promising paradigm for robotic manipulation, yet existing approaches typically support either purely image-based forecasting or reasoning over partial 3D geometry, limiting their ability to predict complete 4D scene dynamics. To solve this, this work explores a novel embodied 4D world model that enables geometrically consistent, arbitrary-view RGBD generation: given only a single-view RGBD observation as input, the model “imagines” the remaining viewpoints, which can then be back-projected and fused to assemble a more complete 3D structure across time. To efficiently learn the multi-view, cross-modality generation, we explicitly design cross-view and cross-modality feature fusion that jointly encourage consistency between RGB and depth and enforce geometric alignment across views. Beyond prediction, converting generated futures into actions is often handled by inverse dynamics, which is ill-posed because multiple actions can explain the same transition. We address this with a test-time action optimization strategy that backpropagates through the generative model to infer a trajectory-level latent best matching the predicted future, and a residual inverse dynamics model that turns this trajectory prior into accurate executable actions. Extensive experiments on the three datasets and platforms demonstrate strong performance on both 4D scene generation and downstream manipulation, and ablations provide practical insights into the key design choices. Project page is available at \url{https://mercerai.github.io/MVISTA-4D/}.

Normalization Equivariance (NE) is a structural prior that improves robustness to distribution shift in image-to-image tasks. A function $f$ is normalization equivariant iff $f(a y + b\mathbf{1}) = a f(y) + b\mathbf{1}$ for all $a>0$ and $b\in\mathbb{R}$. Existing NE methods constrain every internal layer to NE-compatible operations. These constraints add runtime cost and exclude standard transformer components such as softmax attention and LayerNorm. We introduce Wrapped Normalization Equivariance (WNE), a parameter-free wrapper that normalizes the input, applies any backbone, and denormalizes the output. We prove every NE function admits this factorization, so the wrapper exactly parameterizes the class of NE functions. On blind denoising, wrapping CNN and transformer architectures improves robustness under noise-level mismatch with no measurable GPU overhead, while architectural NE baselines are up to $1.6\times$ slower.

Joint-Embedding Predictive Architectures (JEPA) learn view-invariant representations and admit projection-based distribution matching for collapse prevention. Existing approaches regularize representations towards isotropic Gaussian distributions, but inherently favor dense representations and fail to capture the key property of sparsity observed in efficient representations. We introduce Rectified Distribution Matching Regularization (RDMReg), a sliced two-sample distribution-matching loss that aligns representations to a Rectified Generalized Gaussian (RGG) distribution. RGG enables explicit control over expected $\ell_0$ norm through rectification, while its continuous truncated component admits a maximum-entropy characterization under expected $\ell_p$ norm and support constraints. Equipping JEPAs with RDMReg yields Rectified LpJEPA, which strictly generalizes prior Gaussian-based JEPAs. Empirically, Rectified LpJEPA learns sparse, non-negative representations with favorable sparsity--performance trade-offs and competitive downstream performance on image classification benchmarks, showing that RDMReg can enforce sparsity while preserving task-relevant information.

Agentic reinforcement learning has advanced large language models (LLMs) to reason through long chain-of-thought trajectories while interleaving external tool use. Existing approaches assume a fixed inventory of tools, which limits the adaptability of LLM agents to new or evolving toolsets. We present AutoTool, a training framework that equips LLM agents with dynamic tool-selection capabilities throughout their reasoning trajectories. AutoTool employs a dual-phase optimization pipeline: (i) SFT and RL-based trajectory stabilization for coherent reasoning, and (ii) KL-regularized Plackett–Luce Ranking to refine consistent multi-step tool selection. We further build a 200k dataset with explicit tool-selection rationales across 1,000+ tools and 100+ tasks spanning mathematics, science, code generation, and multimodal reasoning. Across ten diverse benchmarks, we train two base models, Qwen3-8B and Qwen2.5-VL-7B, with AutoTool. With fewer parameters, AutoTool consistently outperforms advanced LLM agents and tool-integration methods, yielding average gains of 6.4\% in math \& science reasoning, 4.5\% in search-based QA, 7.7\% in code generation, and 6.9\% in multimodal understanding. In addition, AutoTool exhibits stronger generalization by dynamically leveraging unseen tools from evolving toolsets during inference.

Pretraining optimizers are tuned to produce the strongest possible base model, on the assumption that a stronger starting point yields a stronger model after subsequent changes like post-training and quantization. This overlooks the geometry of the base model which controls how much of the base model's capabilities survive subsequent parameter updates. We study three pretraining optimization approaches that bias optimization toward flatter minima: Sharpness-Aware Minimization (SAM), large learning rates, and shortened learning rate annealing periods. Across model sizes ranging from 20M to 150M parameters, we find that these interventions consistently improve downstream performance after post-training on five common datasets with up to 80\% less forgetting. These principles hold at scale: a short SAM mid-training phase applied to an existing OLMo-2-1B checkpoint reduces forgetting by 31\% after MetaMath post-training and by 40\% after 4-bit quantization.