ICML 2026 Papers

The pursuit of reducing the memory footprint of the self-attention mechanism in multi-headed self attention (MHA) spawned a rich portfolio of methods, e.g., group-query attention (GQA) and multi-head latent attention (MLA). The methods leverage specialized low-rank factorizations across embedding dimensions or attention heads. From the point of view of classical low-rank approximation, these methods are unconventional and raise questions of which objects they really approximate and how to interpret the low-rank behavior of the resulting representations. To answer these questions, this work proposes a generalized view on the weight objects in the self-attention layer and a factorization strategy, which allows us to construct a parameter efficient scheme, called Tucker Attention. Tucker Attention requires an order of magnitude fewer parameters for comparable validation metrics, compared to GQA and MLA, as evaluated in LLM and ViT test cases. Additionally, Tucker Attention~encompasses GQA, MLA, MHA as special cases and is fully compatible with flash-attention and rotary position embeddings (RoPE). This generalization strategy yields insights of the actual ranks achieved by MHA, GQA, and MLA, and further enables simplifications for MLA.

Real-world degradations, such as noise, blur, and low resolution, significantly impair the performance of Multi-modal Large Language Models (MLLMs) in document understanding tasks. Despite recent advancements, progress in this field remains stifled by two critical bottlenecks: the scarcity of large-scale, aligned training data necessary for learning robustness, and the lack of transferable restoration solutions across diverse MLLM architectures. To bridge the data gap, we first present DocRobust-VQA, a large-scale dataset explicitly constructed to support robustness training. Comprising 189K aligned clean/corrupted document image pairs and 417K QA pairs, it provides the first substantial corpus for fine-tuning MLLMs to handle varying degradation conditions. Leveraging this data, we propose Uni-DocRobust, a universal plug-and-play framework that decouples restoration capabilities from specific visual encoders. Our method employs a frozen Universal Restoration Core pre-trained in a canonical feature space via multi-teacher distillation, which can be seamlessly integrated into target MLLMs (e.g., Qwen-VL, InternVL) through lightweight Feature Adapters. Extensive experiments demonstrate that Uni-DocRobust significantly enhances robust performance on MLLMs and enables a cost-effective ``pre-train once, deploy everywhere'' paradigm for robust MLLM deployment.

Neural networks are famously nonlinear. However, linearity is defined relative to a pair of vector spaces, $f:\mathcal{X}\to\mathcal{Y}$. Leveraging the algebraic concept of transport of structure, we propose a method to explicitly identify non-standard vector spaces where a neural network acts as a linear operator. When sandwiching a linear operator $A$ between two invertible neural networks, $f(x)=g_y^{-1}(A g_x(x))$, the corresponding vector spaces $\mathcal{X}$ and $\mathcal{Y}$ are induced by newly defined addition and scaling actions derived from $g_x$ and $g_y$. We term this kind of architecture a Linearizer. This framework makes the entire arsenal of linear algebra, including SVD, pseudo-inverse, orthogonal projection and more, applicable to nonlinear mappings. Furthermore, we show that the composition of two Linearizers that share a neural network is also a Linearizer. We leverage this property and demonstrate that training diffusion models using our architecture makes the hundreds of sampling steps collapse into a single step. We further utilize our framework to enforce idempotency (i.e.\ $f(f(x))=f(x)$) on networks leading to a globally projective generative model and to demonstrate modular style transfer.

High-order relations involving multiple interacting entities are commonly encountered, particularly in biomedical domains. Existing relational learning methods typically learn static entity representations and assume symmetric relation inference, which can be inadequate for capturing context-dependent entity functions and the inherent asymmetry of high-order relations. In this paper, we propose Contextualized High-order Contrastive Learning (CONTEXTOR), a general and plug-and-play framework that formulates high-order relation inference as a dynamic query–response process. Specifically, CONTEXTOR decomposes each high-order relation into multiple incomplete query tuples and their corresponding response entities. Given a query tuple, we contextualize candidate response entity representations via an asymmetric conditional modulation, and align queries with their corresponding contextualized responses through multi-fold contrastive learning. Extensive experiments on benchmark datasets spanning multiple biomedical tasks demonstrate that CONTEXTOR consistently outperforms existing methods across diverse evaluation settings. Code is available at https://github.com/ZJUDataIntelligence/CONTEXTOR.

Fast Healthcare Interoperability Resources (FHIR) is the dominant standard for interoperable exchange of healthcare data. In FHIR, electronic health records form a directed graph of resources. Answering clinically meaningful questions over FHIR requires agents to perform multi-step reasoning, filtering, and aggregation across multiple resource types. Prior work shows that even tool-augmented LLM agents (retrieval, code execution, multi-turn planning) often select the wrong resources or violate traversal constraints. We study this problem in the context of FHIR-AgentBench, a benchmark for realistic question answering over real-world hospital data, and frame reasoning on FHIR as a sequential decision-making problem over a queryable structured graph. We implement a multi-turn CodeAct agent and post-train it with reinforcement learning using a custom harness and tools. A LLM Judge provides execution-grounded rewards. Compared to prompt-based, closed-model baselines, RL post-training improves performance while enforcing data-integrity constraints. Empirically, our approach improves answer correctness from 50% (o4-mini) to 77% on FHIR-AgentBench using a smaller and cheaper Qwen3-8B model. We present an end-to-end post-training pipeline (environment building, harness construction, model training and custom evaluation) that reliably improves multi-turn reasoning over structured clinical graphs.

Federated learning increasingly faces deletion requests that require client-level unlearning without sacrificing model quality, yet a client’s influence is often deeply entangled after many rounds of aggregation. We aim to make unlearning fast, stable, and predictable by reducing the gap to leave-one-out retraining under realistic heterogeneity. We propose Influence-Disentangled Federated Training (IDFT), which instruments standard FedAvg with training-time influence logging: each round’s updates are decomposed into shared covarying directions and a client-separable residual trace, and an entanglement-aware shrinkage suppresses non-removable components. Deletion then becomes a single subtraction followed by a short anchored repair, yielding a stability-style characterization of retrain fidelity driven by the unremoved residual. Across representative benchmarks, IDFT consistently attains the lowest retrain gap (Avg. Gap) on multiple dataset--architecture pairs and improves the fidelity--cost frontier, matching retrain-level forgetting with substantially lower communication/compute than history-heavy baselines. These results suggest a practical pathway to unlearning-friendly federated systems by designing for removability during training rather than relying solely on post-hoc corrections.

Federated Learning (FL) enables privacy-preserving collaboration among distributed clients in open-world environments, but its performance often degrades under data heterogeneity and unpredictable distribution shifts. Test-Time Adaptation (TTA) has recently been introduced into FL to leverage unlabeled data from unseen clients for online adaptation. However, most existing federated TTA methods employ local feature statistics, which can be brittle under diverse and severe distribution shifts. In this work, we observe that despite significant variations in feature distributions, the relational structure among class prototypes—termed *prototype affinity topology*—remains remarkably stable across heterogeneous clients. Building on this insight, we propose FedPAT, a **Fed**erated TTA framework that leverages **P**rototype **A**ffinity **T**opology as a cross-client structural prior. FedPAT learns a global PAT by aggregating class prototypes from source clients, capturing consensus inter-class relationships that are robust to local distribution variations. For unseen target clients, we design a topology-aware mechanism that enhances predictions via diffusion of the global PAT, fuses them with parametric outputs, and performs lightweight optimization for robust test-time adaptation. Extensive experiments demonstrate that FedPAT consistently outperforms advanced federated TTA and classical TTA methods across various distribution shifts.

Quantum error correction (QEC) is essential for enabling quantum advantages, with decoding as a central algorithmic primitive. Owing to its importance and intrinsic difficulty, substantial effort has been made to QEC decoder design, among which neural decoders have recently emerged as a promising data-driven paradigm. Despite this progress, practical deployment remains hindered by a fundamental accuracy–latency tradeoff, often on the microsecond timescale. To address this challenge, here we revisit neural decoders for surface-code decoding under explicit accuracy–latency constraints, considering code distances up to $d=9$ (161 physical qubits). We unify and redesign representative neural decoders into five architectural paradigms and develop an end-to-end compression pipeline to evaluate their deployability and performance on FPGA hardware. Through systematic experiments, we reveal several previously underexplored insights: (i) near-term decoding performance is driven more by data scale than architectural complexity; (ii) appropriate inductive bias is essential for achieving high decoding accuracy; and (iii) INT4 quantization is a prerequisite for meeting microsecond-scale latency requirements on FPGAs. Together, these findings provide concrete guidance toward scalable and real-time neural QEC decoding.

We study strong universal Bayes-consistency in the realizable setting for learning with general metric losses, extending classical characterizations beyond $0$-$1$ classification (Bousquet et al., 2021; Hanneke et al., 2021) and real-valued regression (Attias et al., 2024). Given an instance space $(X,\rho)$, a label space $(Y,\ell)$ with possibly unbounded loss, and a hypothesis class $H \subseteq Y^{X}$, we resolve the realizable case of an open problem presented in Tsir Cohen and Kontorovich (2022). Specifically, we find the necessary and sufficient conditions on the hypothesis class $H$ under which there exists a distribution-free learning rule whose risk converges almost surely to the best-in-class risk (which is zero) for every realizable data-generating distribution. Our main contribution is this sharp characterization in terms of a combinatorial obstruction: Similarly to Attias et al. (2023), we introduce the notion of an infinite non-decreasing $(\gamma_k)$-Littlestone tree, where $\gamma_k \to \infty$. This extends the Littlestone tree structure used in Bousquet et al. (2021) to the metric loss setting.

Group Relative Policy Optimization (GRPO) effectively scales LLM reasoning but incurs prohibitive computational costs due to its extensive group-based sampling requirement. While recent selective data utilization methods can mitigate this overhead, they could induce estimation bias by altering the underlying sampling distribution, compromising theoretical rigor and convergence behavior. To address this limitation, we propose Dynamic Pruning Policy Optimization (DPPO), a framework that enables dynamic pruning while preserving unbiased gradient estimation through importance sampling-based correction. By incorporating mathematically derived rescaling factors, DPPO significantly accelerates GRPO training without altering the optimization objective of the full-batch baseline. Furthermore, to mitigate the data sparsity induced by pruning, we introduce Dense Prompt Packing, a window-based greedy strategy that maximizes valid token density and hardware utilization. Extensive experiments demonstrate that DPPO consistently accelerates training across diverse models and benchmarks. For instance, on Qwen3-4B trained on MATH, DPPO achieves 2.37$\times$ training speedup and outperforms GRPO by 3.36\% in average accuracy across six mathematical reasoning benchmarks.

Multi-objective discrete optimization problems, such as molecular design, pose significant challenges due to their vast and unstructured combinatorial spaces. Traditional evolutionary algorithms often get trapped in local optima, while expert knowledge can provide crucial guidance for accelerating convergence. Large language models (LLMs) offer powerful priors and reasoning ability, making them natural optimizers when expert knowledge matters. However, closed-source LLMs, though strong in exploration, cannot update their parameters and thus cannot internalize experience. Conversely, smaller open models can be continually fine-tuned but lack broad knowledge and reasoning strength. We introduce Multi-LLM Collaborative Co-evolution (MCCE), a hybrid framework that unites a frozen closed-source LLM with a lightweight trainable model. The system maintains a trajectory memory of past search processes; the small model is progressively refined via reinforcement learning, with the two models jointly supporting and complementing each other in global exploration. Unlike model distillation, this process enhances the capabilities of both models through mutual inspiration. Experiments on multi-objective drug design benchmarks show that MCCE achieves state-of-the-art Pareto front quality and consistently outperforms baselines. These results highlight a new paradigm for enabling continual evolution in hybrid LLM systems, combining knowledge-driven exploration with experience-driven learning. The code of MCCE is available on https://github.com/lzz-z/MCCE

Large-scale vision–language models (VLMs) excel at multimodal reasoning, yet efficiency collapses when vision tokens—often orders of magnitude more than text—dominate compute and memory. Prior token-reduction strategies typically trade off salience (which is prone to position bias and incurs extra computation) against diversity (which can under-cover salient regions and is sensitive to hyperparameters). We present SPLIT, a theoretically grounded framework that jointly preserves salience and diversity while aggressively eliminating redundancy. SPLIT (i) estimates token importance via temporal shifts of hidden states across layers—eschewing attention scores and their biases; (ii) assigns adaptive region-level budgets to guarantee localized coverage; and (iii) selects tokens using a diversity score that prioritizes distinctive, non-redundant representations. Our analysis shows that adaptive budgeting yields tighter coverage guarantees than uniform allocation, and our selection rule maintains diversity without costly tuning. Empirically, SPLIT consistently outperforms state-of-the-art on image and video understanding benchmarks. On image understanding with LLaVA-1.5-7B, SPLIT preserves over 99\% accuracy with 192 vision tokens and about 92.8\% with only 64 tokens, demonstrating robust performance under severe token budgets. These results indicate that SPLIT delivers scalable, attention-score-free token reduction that makes multimodal reasoning substantially more efficient without sacrificing accuracy.

Infrared object detection is crucial for perception in autonomous driving and surveillance but remains vulnerable to physical adversarial attacks. Unlike in the RGB domain, where attacks rely on color texture, infrared attacks must manipulate thermal signatures, making the geometry shape of heat-blocking materials the primary adversarial information carrier. Current shape-based methods suffer from a fundamental trade-off between representational capability and optimization power, limiting their attack effectiveness. In this work, we overcome this dilemma by introducing learnable Fourier shapes to the infrared domain. We utilize an end-to-end differentiable framework where a compact set of Fourier coefficients, defining the shape boundary, is analytically mapped to a pixel-space mask via the winding number theorem. This enables efficient gradient-based optimization to generate potent shapes that cause human targets to evade detection. Extensive digital and physical experiments provide a comprehensive evaluation and validate our superior performance. Our resulting physical patch achieves striking robustness, successfully evading detectors across diverse distances, angles, poses, and individuals, and achieves over 88% attack success rate at distances greater than 25m (conf.=0.5). Code is available at https://github.com/Yongyx99/Fourier-shape-attack.

Geometric properties can be leveraged to stabilize and speed reinforcement learning. Existing examples include encoding symmetry structure, geometry-aware data augmentation, and enforcing structural restrictions. In this paper, we take a novel view of RL through the lens of order theory and recast value function estimates into learning a desired poset (partially ordered set). We propose \emph{GCR-RL} (Geometric Coherence Regularized Reinforcement Learning) that computes a sequence of super-poset refinements -- by refining posets in previous steps and learning additional order relationships from temporal difference signals -- thus ensuring geometric coherence across the sequence of posets underpinning the learned value functions. Two novel algorithms by Q-learning and by actor--critic are developed to efficiently realize these super-poset refinements. Their theoretical properties and convergence rates are analyzed. We empirically evaluate GCR-RL in a range of tasks and demonstrate significant improvements in sample efficiency and stable performance over strong baselines.

By synergizing graph topology with the global expressive power of attention, Graph Transformers (GTs) have emerged as a powerful architecture for node classification. Existing GTs mainly focus on diverse topology injection paradigms, which fundamentally construct different propagation operators. However, a unified theoretical understanding of what constitutes a desirable propagation operator remains largely unexplored. To bridge this gap, this paper rethinks GTs from a graph signal denoising perspective, revealing that a block-diagonal structure is a desirable structural prior for graph denoising. To efficiently instantiate this prior, this paper introduces BDFormer, which imposes spectral block regularization on the affinity of a compact set of latent anchors. Furthermore, by introducing hard assignments between nodes and anchors, BDFormer establishes sparse, structured propagation that suppresses noisy cross-class propagation in linear time. Meanwhile, the learned global affinity guides the pruning of inter-class edges in the graph topology, enabling both global and local propagation to jointly adhere to the target block-diagonal structure. Extensive experiments on benchmark datasets demonstrate the superiority and scalability of BDFormer.

Scaling inference-time computation has enabled Large Language Models (LLMs) to achieve strong reasoning performance, but their inherently sequential decoding incurs substantial latency, motivating parallelization of the generation process. However, existing parallel reasoning approaches suffer from performance degradation compared to their sequential counterparts, and often rely on specialized inference engines. We introduce ThreadWeaver, a framework for adaptive parallel reasoning that matches the accuracy of comparably sized sequential reasoning models while significantly reducing inference latency via three key innovations: 1) a two-stage parallel trajectory generator that produces high-quality parallel chain-of-thought data for supervised fine-tuning; 2) a trie-based rollout design that enables parallel reasoning on any off-the-shelf autoregressive inference engine; and 3) a parallelization-aware reinforcement learning framework that trains the model to balance reasoning accuracy with effective parallelization. Across six challenging math reasoning benchmarks, ThreadWeaver trained on top of Qwen3-8B achieves performance on par with cutting-edge sequential reasoning models (79.9% on AIME24 and 71.9% on average) while delivering up to 1.53x speedup in token latency, establishing a new Pareto frontier between accuracy and efficiency.

Influence functions are a standard tool for data debugging and unlearning, but they become impractical for **high-query** subset workloads such as large-$K$ cross-validation, repeated resampling, or interactive what-if analysis as each subset query typically requires an expensive inverse-curvature solve. We introduce **Clustered Influence Functions (CiF)**, which turns subset influence into an **amortized subset oracle**. We build a compact cache once by clustering training gradients, solve a damped Generalised Gauss-Newton system only for cluster means, and answer new subset queries by a linear recombination using cluster membership counts. This yields per-query cost of $O(Cp)$ linear in the cache size $C$, and the number of model parameters $p$. We further provide a diagnostic error bound that decomposes approximation error into a **clustering scatter** term and a **solver residual** term, making the accuracy-compute tradeoff explicit through the cache budget and solver tolerance. Evaluations across MNIST, CIFAR-10 show that CiF matches per-query influence rankings while significantly reducing the total runtime in high-$Q$ regimes, enabling influence-based workflows that are otherwise computationally prohibitive.

Reinforcement learning (RL) is the method of choice for training models in setups where the objective function can only be evaluated by sampling from the model. Our key observation is that when the feedback is terminal and binary, models implicitly induce a likelihood over correct rollouts. Maximum likelihood would be the natural framework in such settings, but RL is used instead as a workaround to the non-differentiability. We prove that the standard, expected-reward RL formulation is only a first-order approximation of the likelihood. To remedy this mismatch, we introduce **Maximum Likelihood Reinforcement Learning (MaxRL)**, a compute-indexed family of sample-based objectives that interpolate between expected-reward RL and maximum likelihood as sampling compute is scaled. The resulting objective is a one-line change to standard RL implementations. MaxRL Pareto-dominates existing methods in all tested models and tasks, achieves up to $\mathbf{20\times}$ gains in test-time scaling efficiency over GRPO, and scales more favorably with additional training data and compute.

Multimodal visual object tracking can be divided into to several kinds of tasks (e.g. RGB and RGB+X tracking), based on the input modality. Existing methods often train separate models for each modality or rely on pretrained models to adapt to new modalities, which limits efficiency, scalability, and usability. Thus, we introduce OneTrackerV2, a unified multi-modal tracking framework that enables end-to-end training for any modality. We propose Meta Merger to embed multi-modal information into a unified space, allowing flexible modality fusion and robustness. We further introduce Dual Mixture-of-Experts (DMoE): T-MoE models spatiotemporal relations for tracking, while M-MoE embeds multi-modal knowledge, disentangling cross-modal dependencies and reducing feature conflicts. With a shared architecture, unified parameters, and a single end-to-end training, OneTrackerV2 achieves state-of-the-art performance across five RGB and RGB+X tracking tasks and 12 benchmarks, while maintaining high inference efficiency. Notably, even after model compression, OneTrackerV2 retains strong performance. Moreover, OneTrackerV2 demonstrates remarkable robustness under modality-missing scenarios.