ICML 2026 Papers

We study how Large Language Models (LLMs) process negation mechanistically. First, we establish that even though open-weight models often provide wrong answers to questions involving negation, they do possess internal components that process negation correctly. Their poor accuracy is due to late-layer attention behavior that promotes simple shortcuts; ablating those attention modules greatly improves accuracy on negation-related questions. Second, we uncover how models process negation. We consider two hypotheses: models could use attention heads that attend to the phrase being negated and suppress related concepts, or they could directly construct a representation of the entire negative phrase (e.g., representing "not gas" as a vector that promotes liquids and solids). We apply a range of observational and causal interpretability techniques on Mistral-7B and Llama-3.1-8B to show that models implement both mechanisms, with the "constructive" mechanism being more prominent. Combined, our work deepens the understanding of LLMs' internals, highlighting construction-dominant computations and the coexistence of competing mechanisms within LLMs.

Large language models usually face continuous domain shifts during testing, which degrade performance on unseen shifting domains. So, researchers propose continual test-time adaptation (CTTA) to adapt to evolving testing domains while preserving knowledge of previous domains, making adaptability-stability (A-S) balance. Existing CTTA methods are constrained by dense base models that encode knowledge from all domains into a global model, hardly achieving the A-S balance. We observe that the model sparsity of mixture-of-experts (MoE) models is better for achieving A–S balance than dense models. In CTTA, however, MoE faces difficulty in (1) correctly routing samples from unseen shifting domains and (2) capturing domain-level shifts. In this paper, we propose test-time mixture-of-experts routing (TiME) via asymmetric co-optimal transport (As-COOT): we model MoE routing in CTTA as a test-time allocation problem via COOT. To ensure reliable routing, we propose a semantic space alignment to align sample-expert distributions via bidirectional contrastive learning. To address COOT’s limitations in CTTA, we propose As-COOT, relaxing sample-side constraints while enforcing expert-side constraints to ensure noise robustness and balance expert load. Experiments show TiME outperforms baselines.

Large-scale learning systems often face the challenge of balancing multiple, potentially competing objectives, such as fairness, accuracy, and latency. While recent work has formalized this as an optimization problem over binary states, many real-world control parameters---such as fairness thresholds, diversity mixing rates, or resource budgets---are continuous. In this work, we extend the framework to *continuous state spaces*. We model the problem as minimizing a sum of linear objectives subject to *movement costs* that penalize system instability. We capture the local structure of the objectives using a *dependency graph* (or factor graph), where each objective is determined by a subset of the state attributes. To address the tension between exploration and stability, we propose *Lazy Graph-LinUCB*, an algorithm that performs lazy updates to minimize switching costs while maintaining near-optimal regret. Beyond stability, we introduce three advanced mechanisms to exploit the underlying graph structure: (1) an *asynchronous* update schedule that eliminates synchronization overhead in sparse graphs; (2) an *adaptive* algorithm that learns the graph structure from data; and (3) a *joint estimator* that leverages data sharing among correlated objectives to significantly tighten regret bounds. Empirically, we demonstrate that these structural exploitations reduce movement costs by more than a factor of three in heterogeneous systems while maintaining similar cumulative losses.

Foundation models have revolutionized AI, but adapting them efficiently for multimodal tasks, particularly in dual-stream architectures composed of unimodal encoders, such as DINO and BERT, remains a significant challenge. Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) enable lightweight adaptation, yet they operate in isolation within each modality, limiting their ability in capturing cross-modal interactions. In this paper, we take a step in bridging this gap with Cross-Modal Low-Rank Adaptation (CoLA), a novel PEFT framework that extends LoRA by introducing a dedicated inter-modal adaptation pathway alongside the standard intra-modal one. This dual-path design enables CoLA to adapt unimodal foundation models to multimodal tasks effectively, without interference between modality-specific and cross-modal learning. We evaluate CoLA across a range of vision-language (RefCOCO, RefCOCO+, RefCOCOg) and audio-visual (AVE, AVS) benchmarks, where it consistently outperforms LORA, achieving a relative gain of around 3\% and 2\%, respectively, while maintaining parameter efficiency. Notably, CoLA enables the first multi-task PEFT framework for visual grounding, bridging a key gap in efficient multimodal adaptation. Code is available at https://github.com/peterwisu/CoLA

Understanding how cellular morphology, gene expression, and spatial context jointly shape tissue function is a central challenge in biology. Image-based spatial transcriptomics technologies now provide high-resolution measurements of cell images and gene expression profiles, but existing methods typically analyze these modalities in isolation or at limited resolution. We address the problem by introducing SPATIA, a multi-level generative and predictive model that learns unified, spatially aware representations by fusing morphology, gene expression, and spatial context from the cell to the tissue level. SPATIA also incorporates a spatially conditioned generative framework with confidence-aware OT reweighting and morphology-profile alignment for modeling target-state morphology distributions. Specifically, we propose a confidence-aware flow matching objective that reweights weak optimal-transport pairs based on uncertainty. We further apply morphology-profile alignment to encourage biologically meaningful image generation, enabling the modeling of microenvironment-dependent phenotypic transitions. We assembled a multi-scale dataset consisting of 25.9 million cell-gene pairs across 17 tissues. We benchmark SPATIA against 18 models across 12 tasks, spanning categories such as phenotype generation, annotation, clustering, gene imputation, and cross-modal prediction. SPATIA achieves improved performance over state-of-the-art models, improving generative fidelity by 8\% and predictive accuracy by up to 3\%.

The rapid evolution of Large Language Models has catalyzed a surge in scientific idea production, yet this leap has not been accompanied by a matching advance in idea evaluation. The fundamental nature of scientific evaluation needs knowledgeable grounding, collective deliberation, and multi-criteria decision-making. However, existing idea evaluation methods often suffer from narrow knowledge horizons, flattened evaluation dimensions, and the inherent bias in LLM-as-a-Judge. To address these, we regard idea evaluation as a knowledge-grounded, multi-perspective reasoning problem and introduce **InnoEval**, a deep innovation evaluation framework designed to emulate human-level idea assessment. We apply a heterogeneous deep knowledge search engine that retrieves and grounds dynamic evidence from diverse online sources. We further achieve review consensus with an innovation review board containing reviewers with distinct academic backgrounds, enabling a multi-dimensional decoupled evaluation across multiple metrics. We construct comprehensive datasets derived from authoritative peer-reviewed submissions to benchmark InnoEval. Experiments demonstrate that InnoEval can consistently outperform baselines in point-wise, pair-wise, and group-wise evaluation tasks, exhibiting judgment patterns and consensus highly aligned with human experts.

Diffusion large language models (dLLMs) represent a promising alternative to autoregressive LLMs; however, the lack of effective post-training techniques, including reinforcement learning (RL), remains a key challenge for dLLMs, especially for downstream applications. Existing approaches often rely on a sequence-level view that requires biased likelihood approximations. In this work, we propose Amortized Group Relative Policy Optimization (AGRPO), a policy gradient algorithm that leverages the Markovian nature of dLLM generation, optimizing individual denoising steps rather than full sequences. Our approach improves the theoretical alignment between training and inference policies and also admits efficient, unbiased gradient updates via a novel timestep estimation scheme. We demonstrate AGRPO's effectiveness on different math and reasoning tasks, achieving absolute accuracy gains of +59.4\% and +69.7\% on Countdown and Sudoku over the base LLaDA model, exceeding comparable methods such as diffu-GRPO.

Auditory attention decoding (AAD) based on Electroencephalography (EEG) aims to identify the attended speaker in multi-speaker environments. However, existing methods typically overlook the crucial phase information of EEG signals, which limits their ability to distinguish structured neural patterns from random noise in the frequency domain and hinders robust decoding. To address these issues, this paper proposes a Phase-aware Complex Refinement Network (PCRNet) for AAD, which consists of a Temporal Context Calibration (TCC) module and a Dual-Domain Integration (DDI) module. Specifically, the TCC module captures long-range temporal dependencies through multi-scale temporal attention mechanism, while the DDI module employs a phase-guided spectral filtering strategy to dynamically suppress noise-dominated frequencies and refine the real and imaginary components separately. This design enables effective phase recalibration and enhances the discriminability of target features in the complex domain. Experimental results on three public datasets demonstrate that PCRNet outperforms state-of-the-art (SOTA) methods, particularly under challenging ultra-short 0.1-second windows. Code is available at: https://github.com/SunshineGreeny/PCRNet.

This paper introduces Probabilistic Multi-Variable Branching (PMVB), a simple and effective technique for accelerating mixed-integer optimization using data-driven machine learning models. At its core, PMVB employs a multi-variable branching procedure that partitions the feasible region via data-driven hyperplanes and requires only two lines of code to implement. Moreover, PMVB is model-agnostic and compatible with a wide range of machine learning models. Leveraging tools from statistical learning theory, we develop interpretable hyperparameter selection strategies and propose several extensions to further enhance performance. We evaluate PMVB by integrating it into state-of-the-art MIP solvers and conducting experiments on both classical benchmark datasets and real-world instances. The results demonstrate the effectiveness of PMVB in improving MIP-solving efficiency.

As large language models increasingly function as autonomous coding agents, code documentation should be designed not for human readability, but for agent executability — serving as repo-specific playbooks that specify precise behaviors agents can follow. We formulate agent-oriented documentation generation as a black-box optimization problem over the documentation space, where quality is defined solely by downstream code correctness. A central challenge for conventional LLM refinement methods is output coupling—program entities are interdependent, and refining the documentation of one entity can invalidate its callers, resulting in a persistent whack-a-mole phenomenon during inference-time scaling. We propose DocSearch, a dependency-guided bi-level search framework that systematically exploits test-time feedback. The outer level conducts a priority search over the program-entity dependency DAG, enforcing a callee-before-caller refinement order to prevent downstream interference. The inner level performs a beam search over documentation refinements, using diversified error message sampling from self-generated unit tests to better exploit diagnostic signals and escape local optima. On DevEval+, DocSearch achieves 90.7\% solve rate with GPT-4o, outperforming the strongest baseline by 32.6\%. Cross-language experiments further demonstrate that optimized documentation transfers effectively to different target programming languages. Code is available at https://github.com/ccsnow127/docsearch.

Diffusion models have recently emerged as expressive policy representations for online reinforcement learning (RL). However, their iterative generative processes introduce substantial training and inference overhead. To overcome this limitation, we propose to represent policies using MeanFlow models, a class of few-step flow-based generative models, to improve training and inference efficiency over diffusion-based RL approaches. To promote exploration, we optimize MeanFlow policies under the maximum entropy RL framework via soft policy iteration, and address two key challenges specific to MeanFlow policies: action likelihood evaluation and soft policy improvement. Experiments on MuJoCo, DeepMind Control Suite and HumanoidBench benchmarks demonstrate that our method, Mean Flow Policy Optimization (MFPO), achieves performance comparable to or exceeding current diffusion-based baselines while considerably reducing training and inference time. Our code is available at https://github.com/dongxiaoyi-xyz/MFPO.

Recent Multimodal Large Language Models (MLLMs) have demonstrated significant progress in perceiving and reasoning over multimodal inquiries, ushering in a new research era for foundation models. However, vision-language misalignment in MLLMs has emerged as a critical challenge, where the textual responses generated by these models are not factually aligned with the given text-image inputs. Existing efforts to address vision-language misalignment have focused on developing specialized vision-language connectors or leveraging visual instruction tuning from diverse domains. In this paper, we tackle this issue from a fundamental yet unexplored perspective by revisiting the core architecture of MLLMs. Most MLLMs are typically built on decoder-only LLMs consisting of a causal attention mechanism, which *limits the ability of the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text)*. To address this problem a MLLM that unlocks causal attention into our proposed modality-mutual attention (MMA) to enable image tokens to attend to text tokens. This simple yet effective design allows MMA to achieve state-of-the-art performance in 12 multimodal understanding benchmarks (**+6.2\% on average across 3 LLMs backbones**) without introducing additional parameters. Our MMA design is intended to be generic, allowing for applications across various modalities, and scalable to accommodate diverse multimodal scenarios.

A plethora of recent works has shown that for wide, overparameterized neural networks, training with Stochastic Gradient Descent (SGD) often leads to interpolation of the training data without sacrificing generalization performance. A key parameter that is not only closely connected to generalization properties, but is also closely tied to other desiderata such as robustness and resistance to adversarial perturbations is the Lipschitz constant of the neural network. While empirically, the Lipschitz constant has been shown not to increase with network width, theoretical findings only provide bounds with logarithmic growth in the width and only for the random initialization of ReLU-networks. In this work, we close this gap for neural networks with smooth activations by showing that, both at random initialization and throughout lazy training, the local Lipschitz constant of deep neural networks does not increase with network width. More precisely, we establish novel non-asymptotic (finite width) upper bounds and corroborate them by numerical experiments.

We propose a gradient preconditioning method that makes reward-guided generation with one-step generative models both efficient and reliable. Test-time noise optimization can unlock substantially better reward-guided generations from pretrained generative models, but it is prone to reward hacking that degrades quality and is often too slow for practical use. We precondition reward gradients by projecting them onto a carefully designed white Gaussian noise feasible set, a compact spectral set with blockwise norm constraints that tightly captures the statistics and spatial uncorrelatedness of white Gaussian noise. This preconditioning reshapes each gradient update into a noise-aligned direction, driving faster and more effective reward ascent while preventing reward hacking. The projection is closed-form and matches the $\mathcal{O}(N \log N)$ complexity of FFT, adding negligible overhead in practice. In experiments on FLUX with four reward models, our approach reaches a comparable Aesthetic Score using only 30\% of the wall-clock time required by the state-of-the-art regularization-based method.

Multi-domain graph pre-training is a crucial step in constructing foundational graph models with cross-domain generalization capabilities. However, existing methods predominantly rely on jointly training all source domain graphs, resulting in high computational costs. Furthermore, it remains unclear whether all source domain graph data contribute equally to effective transfer. This paper empirically reveals significant data redundancy in multi-domain graph pre-training. Based on this finding, we propose the Multi-domain Graph Pre-training Framework, MDGMIX, which combines boundary-aware subgraph mixing with hierarchical discrimination. By selecting boundary nodes to construct challenging mixed-domain subgraphs, MDGMIX employs coarse-grained domain discrimination and fine-grained domain decomposition losses to decouple shared patterns from domain-specific patterns. During adaptation, MDGMIX employs a lightweight prompt weighting mechanism to transfer source domain knowledge. Extensive experiments demonstrate that MDGMIX consistently outperforms strong baselines in few-shot classification tasks while exhibiting superior time and memory efficiency. The code is available at: https://github.com/zhengziyu77/MDGMIX

Instruction tuning aligns large language models, including multimodal ones, with diverse user intents, but scaling to heterogeneous mixtures is hindered by gradient interference and bandwidth-heavy synchronization. We ask whether these two bottlenecks can be addressed jointly by training parts of the mixture independently and reconciling them once in parameter space. We develop a local quadratic theory inside a shared flat basin that yields three results: weight merging produces a curvature-weighted variance reduction; PCA-aligned conflict splitting maximizes this gain along high-curvature directions; and merging additionally acts as spectral filtering with implicit norm regularization. These results directly motivate **MERIT**, a decentralized merge-ready instruction-tuning pipeline that estimates dataset-level gradient conflicts, partitions the mixture along the top PCA conflict axes, fine-tunes each partition independently with no inter-partition communication, and merges once via token-weighted averaging. On Qwen2.5-VL-3B with 136 Vision-FLAN tasks, MERIT improves the 8-benchmark average from 54.3 (joint training) to 57.0. The same recipe scales to a 7B model on a 1.6M-example, 176-source mixture—matching or exceeding centralized joint training with minimal cost overhead—and transfers to text-only FLAN. Our code is available at [https://github.com/naver-ai/merit](https://github.com/naver-ai/merit).

Reinforcement learning (RL) faces a persistent tension: policies that are stable to optimize (e.g., Gaussians) are often too simple to represent the multimodal action distributions required for complex control. Conversely, expressive generative policies—such as diffusion and flow matching—can be difficult to optimize in online RL due to intractable likelihoods and gradients propagating through long sampling chains. We address this tension with a key structural principle: *decoupling optimization from generation*. Building on this, we introduce GoRL (Generative Online Reinforcement Learning), an algorithm-agnostic framework that trains expressive policies from scratch by confining policy optimization to a tractable latent space while delegating action synthesis to a conditional generative decoder. Using a two-timescale alternating schedule and anchoring decoder refinement to a fixed prior, GoRL enables stable optimization while continuously expanding expressiveness. Empirically, GoRL consistently outperforms unimodal and generative baselines across diverse continuous-control tasks. Notably, GoRL achieves returns exceeding **870** on HopperStand, more than $3\times$ the strongest baseline; on high-dimensional humanoid tasks, it further outperforms the strongest non-GoRL baseline by over an order of magnitude.

Diffusion language models, especially masked discrete diffusion models, have achieved great success recently. While there are some theoretical and primary empirical results showing the advantages of latent reasoning with looped transformers or continuous CoT, continuous diffusion models typically underperform their discrete counterparts. In this paper, we argue that diffusion language models do not necessarily need to be in the discrete space. In particular, we prove that continuous diffusion models have stronger expressivity than discrete diffusions and looped transformers. We attribute the contradiction between the theoretical expressiveness and empirical performance to their practical trainability: while continuous diffusion provides intermediate supervision that looped transformers lack, they are harder to generate and decode tokens in the continuous representation space compared with discrete states. We therefore propose **C**oevolutionary **C**ontinuous **D**iscrete **D**iffusion (CCDD), which defines a joint multimodal diffusion process on the union of a continuous representation space and a discrete token space, leveraging a single model to simultaneously denoise in the joint space. By combining two modalities, CCDD is expressive with rich semantics in the latent space, as well as good trainability and sample quality with the help of explicit discrete tokens. We also propose effective architectures and advanced training/sampling techniques for CCDD, which reveals strong empirical performance in extensive language modeling experiments on real-world tasks.

Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.