ICML 2026 Papers

Recent advances in large-scale diffusion models have intensified concerns about their potential misuse, particularly in generating realistic yet harmful or socially disruptive content. This challenge has spurred growing interest in effective machine unlearning, the process of selectively removing specific knowledge or concepts from a model without compromising its overall generative capabilities. Among various approaches, Low-Rank Adaptation (LoRA) has emerged as an effective and efficient method for fine-tuning models toward targeted unlearning. However, LoRA-based methods often exhibit limited adaptability to concept semantics and struggle to balance removing closely related concepts with maintaining generalization across broader meanings. Moreover, these methods face scalability challenges when multiple concepts must be erased simultaneously. To address these limitations, we introduce UnHype, a framework that incorporates hypernetworks into single- and multi-concept LoRA training. The proposed architecture can be directly plugged into Stable Diffusion as well as modern flow-based text-to-image models, where it demonstrates stable training behavior and effective concept control. During inference, the hypernetwork dynamically generates adaptive LoRA weights based on the CLIP embedding, enabling more context-aware, scalable unlearning. We evaluate UnHype across several challenging tasks, including object erasure, celebrity erasure, and explicit content removal, demonstrating its effectiveness and versatility.

We study bilinear bandits with partially observable features on both the user and item sides. In each of $T$ rounds, the learner selects a user--item pair and observes only the reward for the chosen pair. The reward model is bilinear in the user and item features with an unknown parameter matrix. Existing work commonly reduces this problem to a linear bandit via Kronecker-product features, increasing dimensionality and losing bilinear structure. We propose \texttt{BiRoLF}, an algorithm robust to latent features, which directly leverages the bilinear structure without such linearization. It augments observed feature spaces on both sides with orthogonal complement bases and employs doubly robust (DR) estimation to impute rewards for unselected pairs, constructing matrix-valued pseudo-rewards. We estimate the effective parameter using a bilinear DR-Lasso estimator, which promotes sparsity in components orthogonal to observed features. \texttt{BiRoLF} achieves a $\tilde{O}(\sqrt{(d_x + d_{h_x})(d_y + d_{h_y}) T})$ regret bound, where $d_x$ and $d_y$ are observable feature dimensions, and $d_{h_x}$ and $d_{h_y}$ denote effective orthogonal-complement dimensions. By exploiting the induced block-diagonal Gram structure, \texttt{BiRoLF} performs an exact blockwise update that preserves the full DR-Lasso objective. Numerical experiments show strong regret performance and computational gains.

Large Language Models with Chain-of-Thought reasoning capabilities represent valuable intellectual property, yet existing black-box watermarking methods often trade robustness for reasoning fidelity by perturbing final answers or relying on fragile trigger patterns. We propose BiCoT, a watermarking framework that embeds ownership signals into the internal geometry of reasoning traces by aligning high-saliency structural anchors with a private signature subspace while regularizing ordinary control tokens to preserve semantic capacity. This design couples the watermark with reasoning-relevant representations, making removal difficult without disrupting the features that support coherent reasoning. To enable verification under model theft and representation drift, we introduce Robust Subspace Registration(RSR), a Top-$k$ logprob-based black-box verifier that uses sentinel tokens to calibrate systematic shifts in the output distribution. Experiments show that BiCoT preserves reasoning fidelity across diverse complex reasoning tasks while achieving robust detection under fine-tuning, quantization, model-level perturbations, and adaptive output-level attacks across in-domain and out-of-distribution settings.

Stability is a property of fundamental importance in real-world systems. Although it has been widely studied and well understood in *control theory* (CT) for deterministic systems, it is largely overlooked in stochastic systems such as *Markov decision processes* (MDPs). In this paper, we aim to translate the steady-state control problem, well established in CT, where the goal is to synthesize a controller with prescribed asymptotic stability properties, into the MDP framework. To this end, we propose the novel *fast-mixing steady-state* (FMSS) problem. Given an ergodic MDP and a target steady-state distribution, the objective is to synthesize a Markovian policy that induces this distribution with the fastest possible convergence rate. Addressing this problem requires controlling the spectral properties of the induced *Markov chain* (MC) transition matrix, which generally leads to non-convex programs. Thus, we derive a tractable surrogate objective that leads to a convex program, whose properties we study in terms of approximation quality, feasibility, and computational complexity. We then move to the learning setting and propose an "offline" sample-based algorithm for FMSS (FMSS-SV), designed for tabular MDPs, in which the environment’s transition model is estimated from data. We quantify the impact of transition model estimation errors on both the objective value and the learned policy, and provide a finite-sample complexity analysis.

Mixture-of-Experts (MoE) scales model capacity while preserving computational efficiency through sparse expert activation. However, training high-quality MoEs from scratch is prohibitively expensive. An alternative is to convert pretrained dense models into sparse MoEs. Existing dense-to-MoE methods fall into two categories: \textbf{dynamic structural pruning} that converts dense models into MoEs with moderate sparsity to balance performance and efficiency, and \textbf{downcycling} approaches that use pretrained dense models to initialize highly sparse MoEs. However, existing methods break the intrinsic activation patterns within dense models, leading to suboptimal expert construction. In this work, we argue that the Gated Linear Unit (GLU) provides a natural blueprint for dense-to-MoE conversion. We show that the fine-grained neuron-wise activation patterns of GLU reveal a coarse-grained structure, uncovering an inherent MoE architecture composed of consistently activated universal neurons and dynamically activated specialized neurons. Leveraging this discovery, we introduce ExpertWeaver, a training-free framework that partitions neurons according to their activation patterns and constructs shared experts and specialized routed experts with layer-adaptive configurations. Experiments demonstrate that ExpertWeaver outperforms existing methods, both as a training-free dynamic structural pruning technique and as a downcycling strategy for MoE initialization.

While Large Vision-Language Models (VLMs) excel at interpolation, they suffer catastrophic failures in systematic generalization, most notably in visual counting beyond training distributions. In this work, we investigate this extrapolation bottleneck by deconstructing visual counting into three cognitive stages: object individuation, abstract magnitude representation, and symbolic decoding. Using a controlled environment of synthetic Go game boards, we isolate the specific mechanism of failure. Contrary to the hypothesis that models suffer from perceptual errors, we demonstrate via linear probing that visual backbones maintain robust, linearly separable representations of quantity well into the extrapolation regime. Furthermore, models retain latent magnitude awareness, successfully performing comparative reasoning on quantities they fail to enumerate. We pinpoint the collapse to the Symbolic Decoding stage, where the model fails to project valid visual magnitudes onto discrete tokens. Our findings support a Fractured Magnitude Hypothesis: VLMs fail to acquire a Universal Number Space, instead learning disjoint, modality-specific statistical manifolds that prevent cross-modal grounding for unseen pairings. We validate our findings on the state-of-the-art foundation model, suggesting that bridging the extrapolation gap requires inductive priors that enforce unified magnitude representations rather than simply scaling training data.

CLIP delivers strong zero-shot classification but remains highly vulnerable to adversarial attacks. Prior adversarial fine-tuning work primarily matches predicted logits between clean and adversarial examples, which overlooks uncertainty calibration and may degrade the zero-shot generalization. A common expectation in reliable uncertainty estimation is that predictive uncertainty should increase as inputs become more difficult or shift away from the training distribution. However, we frequently observe the opposite in the adversarial setting: perturbations not only degrade accuracy but also suppress uncertainty, leading to severe miscalibration and over-confidence. This reveals a critical reliability gap beyond robustness. To bridge this gap, we propose an adversarial fine-tuning objective for CLIP considering both accuracy and uncertainty. By reparameterizing CLIP outputs as the concentration parameters of a Dirichlet distribution, we propose a unified representation that captures relative semantic structure and confidence magnitude. This enables holistic distribution alignment under perturbations, moving beyond single-logit anchoring and restoring calibrated uncertainty. Experiments across multiple zero-shot benchmarks demonstrate that our method significantly improves uncertainty calibration and achieves competitive adversarial robustness while preserving clean accuracy.

We study the problem of learning Nash equilibria in offline two-player zero-sum Markov games. While existing approaches often rely on explicit pessimism to address distribution shift, we show that KL regularization alone suffices to stabilize learning and guarantee convergence. We first introduce Regularized Offline Sequential Equilibrium (ROSE), a theoretical framework that achieves a fast $ \widetilde{\mathcal{O}}(1/n) $ convergence rate under *unilateral concentrability*, improving over the standard $ \widetilde{\mathcal{O}}(1/\sqrt{n}) $ rates in unregularized settings. We then propose Sequential Offline Self-play Mirror Descent (SOS-MD), a practical model-free algorithm based on least-squares value estimation and iterative self-play updates. We prove that the last iterate of SOS-MD attains the same $ \widetilde{\mathcal{O}}(1/n) $ statistical rate up to a vanishing optimization error of order $ \widetilde{\mathcal{O}}(1/\sqrt{T}) $ in the number of self-play iterations $T$.

Modern optimizers combine gradients from the current mini-batch with historical optimization state, such as momentum or adaptive moments. While highly effective, aggregating across the batch and incorporating this history can produce parameter updates that increase the loss of individual samples. We term this effect *harm* and formalize the parameter update as an optimization problem that explicitly minimizes the conflicting impact of both batch averaging and past optimization state on current data. Because the exact formulation is intractable, we introduce a highly efficient proxy. We first reduce the problem's dimensionality to the batch size, and then drastically cut memory and speed bottlenecks by successfully restricting the optimization to the last linear layer. This hinges on the unexpected finding that *this layer alone reliably captures the second-order statistics of the per-sample gradients*. The resulting surrogate problem integrates readily into standard optimizers like SGD and AdamW, and can be solved using a small number of GPU-friendly iterations. Crucially, the method exhibits *favorable scaling properties*, as the relative computational overhead shrinks as the model size or input grows. Experiments on image classification benchmarks confirm *reduced per-sample interference* and *improved generalization*.

Group fairness can ensure equitable performance across different demographic subgroups for medical image analysis. However, the current fine-tuned foundation models (FMs) exhibit significant subgroup disparity. One-shot federated learning (OFL) can potentially mitigate this by leveraging cross-institutional data diversity within a single communication round. However, heterogeneous distributions across medical institutions may cause OFL local models to diverge severely, resulting in parameter conflicts that amplify disparity upon aggregation. To address these challenges, we propose Fair-FedMOE, a group-fair OFL framework for medical FMs. During local training, Fairness-aware Expert Routing leverages learnable prototypes to route samples to group-specific experts, enabling subgroup-specialized learning to capture group-specific features without inter-group interference. During model aggregation, Prototype-guided Differential Aggregation computes personalized weights based on prototype similarity and applies differentiated aggregation strategies to filter conflicting updates. We propose RES-AUC, a Rawlsian justice-inspired metric based on worst-group performance that remains stable as groups increase. Extensive experiments on retinal and chest X-ray datasets with multiple FMs demonstrate consistent fairness gains without sacrificing accuracy.

Preference optimization for diffusion models relies on reward functions that are both discriminative and computationally efficient. Vision-Language Models (VLMs) have emerged as powerful reward providers. However, their computation and memory cost can be substantial, and optimizing a latent diffusion generator through a pixel-space reward introduces a domain mismatch that complicates alignment. In this paper, we propose DiNa-LRM, a diffusion-native latent reward model that formulates preference learning directly on noisy diffusion states. Our method introduces a noise-calibrated Thurstone likelihood with diffusion-noise-dependent uncertainty. DiNa-LRM leverages a pretrained latent diffusion backbone with a timestep-conditioned reward head, and supports inference-time noise ensembling, providing a diffusion-native mechanism for test-time scaling and robust rewarding. Across image alignment benchmarks, DiNa-LRM substantially outperforms existing diffusion-based reward baselines and achieves competitive performance compared to state-of-the-art VLMs while maintaining a substantially lower computational cost. In preference optimization, we demonstrate that DiNa-LRM improves preference optimization dynamics, enabling faster and more resource-efficient model alignment.

Looped Language Models (LoopLMs) perform multi-step latent reasoning prior to token generation and outperform conventional LLMs on reasoning benchmarks at smaller parameter budgets. However, attempts to further improve LoopLM reasoning with reinforcement learning have failed—standard objectives such as Group Relative Policy Optimization (GRPO) only assign credit to the final latent state, creating a fundamental mismatch with the model's internal computation. To resolve this, we introduce RLTT (Reward Latent Thought Trajectories), a reinforcement learning framework which distributes reward across the full latent reasoning trajectory. RLTT provides dense, trajectory-level credit assignment without relying on external verifiers and can directly replace GRPO with negligible overhead. Across extensive experiments with Ouro-1.4B/2.6B-Thinking under identical training and inference conditions, RLTT yields statistically significant improvements over GRPO on challenging mathematical reasoning benchmarks, improving mean accuracy over MATH-500, AIME24/26, and BeyondAIME by +5.8\% on the 1.4B scale, and +10.9\% on the 2.6B scale. Despite being trained exclusively on mathematics, RLTT also transfers effectively to non-mathematical reasoning benchmarks, demonstrating the effectiveness of trajectory-level credit assignment for reinforcement learning in LoopLMs. Code is available at https://github.com/jonwill8/RLTT.git.

Causal inference in spatial domains faces two intertwined challenges: (1) unmeasured spatial factors, such as weather, air pollution, or mobility, that confound treatment and outcome, and (2) interference from nearby treatments that violate standard no-interference assumptions. While existing methods typically address one by assuming away the other, we show they are deeply connected: *interference reveals structure* in the latent confounder. Leveraging this insight, we propose the **Spatial Deconfounder**, a two-stage method that reconstructs a substitute confounder from local treatment vectors using a conditional variational autoencoder (C-VAE) with a spatial prior, then estimates causal effects with a flexible outcome model. We show that this enables nonparametric identification of direct and spillover effects under weak assumptions—without multiple treatment types or a known latent-field model. Empirically, we extend SpaCE, a benchmark suite for spatial confounding, to include treatment interference, and show that the Spatial Deconfounder consistently improves effect estimation across real-world environmental health and social science datasets. By turning local interference into a multi-cause proxy for latent spatial confounding, our framework advances robust causal inference for spatial data.

We study optimization over non-convex constraint sets that are homeomorphic to a ball, encompassing important problem classes such as star-shaped sets that frequently arise in machine learning and engineering applications. We propose **Hom-PGD$^+$**, a learning-based and projection-efficient first-order method that efficiently solves such problems without requiring expensive projection or optimization oracles. Our approach leverages an invertible neural network (INN) to learn the homeomorphism between the non-convex constraint set and a unit ball, transforming the original problem into an equivalent ball-constrained optimization where projections admit efficient solutions. We establish that Hom-PGD$^+$ achieves an $\mathcal{O}(\epsilon^{-2})$ convergence rate to an ($\epsilon + \mathcal{O}(\sqrt{\epsilon_{\text{inn}}})$)-approximate stationary solution, where $\epsilon_{\text{inn}}$ denotes the homeomorphism learning error. This rate significantly improves upon existing methods for optimization over non-convex sets, while maintaining a per-iteration complexity of only $\mathcal{O}(W)$ for $W$ INN parameters. Extensive experiments, including QCQP, chance-constrained power-system optimization, and non-uniform adversarial attacks, demonstrate that Hom-PGD$^+$ achieves competitive solution quality while delivering speedups of up to one order of magnitude.

Large-scale molecular dynamics simulations are essential in understanding chemical and biological processes, necessitating the accurate and efficient modeling of interatomic interactions. Existing learning-based methods are generally based on message passing mechanisms; they are either not scalable or too coarse to offer accurate modeling. We propose a new message passing framework that can effectively and efficiently model interatomic interactions for simulating large-scale molecular dynamics at full atomic resolution. Specifically, our framework is stacked with a sequence of message passing neural network layers, each realizing the message passing over a distinct and dilated star-structured path. These star-structured paths are constructed progressively along dilated regions to capture the distance-dependent interactions. The crux of our framework is that it resolves the problem of dense interatomic interactions of large-scale atomic systems with sparser and region-based message passing graphs. We evaluate the framework on four benchmarks: MD22 (molecules with 42–370 atoms), Chignolin (a 166-atom protein featuring diverse conformations), the AdK dataset (a protein trajectory with up to 3,000 atoms), and the MISATO dataset (over 10,000 heterogeneous protein-ligand complexes with systems up to 40,000 atoms). Comprehensive evaluations demonstrate that our approach delivers state-of-the-art performance overall across various benchmarks. The code is provided on [Github](https://github.com/WanyuGroup/ICML2026-DKMP).

While low-rank decomposition offers potential for reducing LLM parameters, maintaining the original capabilities remains a significant challenge. In this work, we identify and formalize a key overlooked issue in LLM decomposition: \textit{representation drift}. We show that approximation errors introduced by decomposition propagate and amplify non-linearly through the deep layers of the transformer architecture, progressively distorting internal representations and degrading downstream performance. To mitigate this, we introduce a conceptually simple but principled compensation mechanism, named ``Decomper'', that operates by suppressing error at its source. By learning to align the output distribution of decomposed transformer blocks with their original counterparts, our method effectively counteracts representation drift, achieving notable performance recovery with near-zero inference overhead. Extensive experiments in OPT, LLaMA-2/3, and Qwen exhibit remarkable improvements. For instance, on LLaMA-3-8B and OPT-13B at 40\% compression, perplexity is reduced by more than 70\% while reasoning task accuracy improves by over 10\%. Our code is available at this \href{https://github.com/xinhaoH/SVD-Compen}{URL}.

Instruction tuning (IT) is a central mechanism for aligning large language models (LLMs) with user intent. In practice, randomly shuffling the training set is a simple yet surprisingly strong baseline. However, it overlooks latent structure, such as domain and reasoning depth, and thus interleaves heterogeneous objectives, which can induce gradient conflicts and diminish effective optimization progress. To this end, we propose **EP-Order**, an embedding-proximity-based data-ordering paradigm for IT of LLMs. Unlike previous paradigms that derive order from **per-example** scores, EP-Order explicitly accounts for **inter-sample** correlations by operating in representation space. EP-Order trains a warm-up model on a small subset of data (*e.g.*, 10%), embeds all training samples for clustering, and ranks these clusters according to embedding proximity. To handle sharp gradient changes at cluster transitions and alleviate catastrophic forgetting under cluster-based training, we introduce mixed regions that interleave samples from the previous, current, and next clusters, thereby stabilizing learning. Extensive experiments on 14 benchmarks spanning vision-language, text-only, and hybrid thinking/no-thinking scenarios show that EP-Order achieves broadly consistent improvements over random shuffling. These results suggest that exploiting embedding-level data structure offers an effective and general direction for improving IT on complex, high-conflict training data.

Long-term weather forecasting is critical for socioeconomic planning and disaster preparedness. While recent approaches employ finetuning to extend prediction horizons, they remain constrained by the issues of catastrophic forgetting, error accumulation, and high training overhead. To address these limitations, we present a novel pipeline across pretraining, finetuning and forecasting to enhance long‑context modeling while reducing computational overhead. First, we introduce an Efficient Multi‑scale Transformer (EMFormer) to extract multi‑scale features through a single convolution in both training and inference. Based on the new architecture, we further employ an accumulative context finetuning to improve temporal consistency without degrading short‑term accuracy. Additionally, we propose a composite loss that dynamically balances different terms via a sinusoidal weighting, thereby adaptively guiding the optimization trajectory throughout pretraining and finetuning. Experiments show that our approach achieves great performance in weather forecasting and extreme event prediction, substantially improving long-term forecast accuracy. Moreover, EMFormer demonstrates strong generalization on vision benchmarks (ImageNet-1K and ADE20K). Code: https://github.com/chenhao-zju/emformer

Self-supervised learning for physiological time-series aims to captures the identity of the underlying dynamical process while filtering irrelevant noise. However, existing approaches may obscure the clinical semantics important for downstream transferability. Weakly constrained pretext tasks (i.e. contrastive learning, MAE) may incorrectly ignore the underlying dynamical structure, while structurally constrained models (i.e. SVAEs) are unable to selectively filter sample-specific noise. To bridge this gap, we propose ${\bf PULSE}$, a novel pretraining objective that simultaneously preserves dynamical relationships important to physiological time-series while selectively removing irrelevant noise. We achieve this by formulating a dynamical systems model to identify transferable and non-transferable information between time-series windows, and target the former through a novel cross-reconstruction objective. We establish theory that provides conditions for when transferrable information is recovered, and empirically validate it through synthetic experiments. On several real-world datasets, PULSE effectively distinguishes clinical semantic classes, increases label efficiency, and improves transfer learning performance.