ICML 2026 Papers

The scaling of Large Language Models (LLMs) has driven significant performance gains but created substantial challenges in inference efficiency. While Mixture of Experts (MoEs) architectures address this by decoupling model size from inference cost, training MoEs from scratch is often unstable and compute intensive. Conversion of pre-trained dense models into sparse MoEs has emerged as an alternative solution; however, existing methods typically rely on heuristic neuron clustering or random splitting to partition Feed-Forward Networks (FFNs) into experts. In this work, we propose DOT-MoE, a novel framework that formulates the decomposition of dense layers as a Differentiable Optimal Transport (DOT) problem. Instead of static heuristics, we model neuron assignment as a balanced transport problem, utilizing differentiable Sinkhorn-Knopp iterations to enforce strict expert capacity constraints. Furthermore, we utilize Straight-Through Estimators (STE) to jointly learn the discrete neuron-to-expert assignment and the token-to-expert routing policy end-to-end. Extensive experiments across multiple architectures and benchmarks demonstrate that DOT-MoE significantly outperforms structured pruning, heuristic clustering, and random-split baselines, retaining 90% of the original dense model's performance while reducing active parameters by 50%.

Heterogeneous federated learning (HtFL) has emerged as a promising approach to address heterogeneity in local computational resources and data distribution. However, existing methods cause performance degradation of model personalization because personalized and generalized knowledge are either intertwined or dominated by one of them. To address this issue, we propose a novel Elastic Mixture of Experts (EMoE) architecture on HtFL, namely FedEMoE, decoupling personalization from generalization. Specially, FedEMoE employs a multi-scale feature extraction mechanism via personalized experts to enrich personalized knowledge. Furthermore, we design an elastic shared expert to break the transferred knowledge bottleneck across heterogeneous client models. The elastic shared expert can adaptively expand or shrink according to the status of each expert by the weight spectrum analysis, respectively. Extensive experiments across statistical and model heterogeneity settings demonstrate that FedEMoE significantly outperforms state-of-the art methods on the accuracy of each heterogeneous model over diverse datasets.

Adapting large-scale pre-trained video generators for Video Super-Resolution (VSR) in novel domains remains computationally prohibitive. Methods that reformulate generation as direct Low-Quality to High-Quality mappings deviate from the original generative formulation, demanding extensive fine-tuning. ControlNet-style adapters lose their efficiency under modern Diffusion Transformers since the absence of encoder-decoder hierarchy forces duplication of the entire backbone. We observe that flow matching offers a principled alternative for cross-domain VSR adaptation. By predicting a constant velocity field across all timesteps, the adaptation task reduces to learning a fixed injection pattern rather than time-varying transformations. Building on this insight, we propose LiteVSR, a minimalist framework that performs VSR using a completely frozen Diffusion Transformer with a lightweight State-Aware Adapter. The adapter employs a dual-stream architecture that extracts static structural cues from the LQ input and dynamic cues from intermediate denoising states, aligning them through time-dependent cross-attention to enable adaptive transition from structural alignment to texture refinement as denoising proceeds. LiteVSR achieves competitive restoration quality with only 11.25% trainable parameters and 12 GPU-hours of training on a single A100, while maintaining fast sampling (down to a single step) compatibility.

Existing decoding-time safety interventions are often reactive, relying on local signals to correct unsafe outputs after they emerge. Under adversarial prompts that drive generation into recurring unsafe response, such local signals provide weak guidance for stable repair. As a result, rollback and post-hoc rewriting often trade-off response quality with recurrent violations. To address these limitations, we propose RBCBF, a rollback-based decoding-time framework that jointly selects intervention steps and performs distribution-level corrective control. Our key innovation is a risk-aggregation formulation that views terminal violations as the accumulated build-up of risk along the prefix. By selecting rollback steps from these decisive prefixes, RBCBF moves rollback targeting beyond heuristic cues and turns it into a trajectory-level decision. RBCBF then applies invasive corrective control to the next-token distribution under multiple rule constraints. Across jailbreak-style evaluations, RBCBF outperforms prior rollback methods and decoding-time baselines, reducing harmful responses and substantially lowering violation recurrence.

Submodular maximization subject to a knapsack constraint (SMK) is a fundamental problem in discrete optimization, with wide-ranging applications in machine learning and related fields. As these applications increasingly involve sensitive individual data, there is a growing need for high-utility algorithms that provide formal privacy guarantees. In this work, we study the SMK problem under differential privacy, considering both monotone and non-monotone objective functions. For monotone objectives, we propose a differentially private algorithm that achieves the optimal $(1-1/e)$-approximation ratio while significantly improving both additive error and query complexity over prior work. We also present a more efficient algorithm for the same setting, achieving a $1/2$-approximation. For non-monotone objectives, we introduce, to our knowledge, the first differentially private algorithm with provable guarantees, achieving a $1/4$-approximation in expectation and an additive error comparable to the best known for monotone objective functions.

Two-server secure inference allows a client to query a hosted large language model (LLM) without revealing prompts or embeddings. Recent GPU systems based on function secret sharing (FSS) make linear layers efficient, but fixed-point nonlinearities and helper operations remain a bottleneck because each operator is typically implemented as a bespoke protocol with its own comparisons, wrap-around corrections, and preprocessing material. We present FuseFSS, a compiler that replaces per-operator protocol design with a single compilation pipeline. For each scalar fixed-point operator, a compact specification lists its interval partition, low-degree arithmetic pieces, and required predicate bits. The compiler emits two batched FSS evaluations on the public masked value: one packed comparison that returns all predicate bits, and one vector interval lookup that returns the active coefficients and constants. Compared to the current state-of-the-art FSS-based GPU secure inference, FuseFSS preserves accuracy while achieving a $1.24\times$–$1.50\times$ end-to-end speedup and reducing online communication by 9%–16% on BERT and GPT-style models; preprocessing is also lighter, with 14%–23% lower key-generation time and 20%–24% smaller keys.

In recent years, significant advancements in large language models have greatly propelled the development of Text-to-SQL tasks. However, due to the token-by-token sequential generation mechanism employed by these models, they encounter a semantic blind spot problem with respect to pending SQL components—the parts of the SQL query yet to be generated. Specifically, language models are unable to effectively utilize the semantic information of these pending SQL components during the generation of the final SQL query, which poses considerable challenges for generating complex SQL statements. To address this issue, we propose a novel thought process based on SQL components pre-generation and design a maximum connected subtree matching reward mechanism leveraging the SQL abstract syntax tree to improve the accuracy of local component generation. Extensive experiments demonstrate that, under comparable model parameter scales, our training approach achieves significant advantages, effectively enhancing the generation of complex SQL queries. Our method attains an execution accuracy EX of 65.78% on the BIRD-dev dataset and achieves state-of-the-art performance on the Spider-syn datasets.

Accurate weather forecast reporting enables individuals and communities to better plan daily activities, agricultural operations, and transportation. However, the current reporting process primarily relies on manual analysis of multi-source data, which often leads to information overload and reduced efficiency. With the rapid advancement of multimodal large language models (MLLMs), leveraging data-driven models to analyze and generate reports in the weather forecasting domain remains largely underexplored. In this work, we propose the Weather Forecasting Report (WFR) task and construct the first instruction-tuning dataset for this task, named WSInstruct, which covers 31 cities in America and 8 weather aspects. Based on this corpus, we develop the first model, WeatherSyn, specialized in generating weather forecast reports. Evaluation across multiple metrics on our dataset shows that WeatherSyn consistently outperforms leading closed-source MLLMs, particularly on structurally complex weather aspects. We further analyze its performance across diverse geographic regions and weather aspects. WeatherSyn demonstrates strong transferability across different regions, highlighting its zero-shot generalization capability. WeatherSyn offers valuable insight for developing MLLMs specialized in weather report generation.

In many prediction problems, we have extra information during training (for example, measurements that are expensive or slow to collect) that will not be available when the model is deployed. A common strategy is to first train a model that uses all training information, then use its predictions on unlabeled examples to train a second model that only uses the inputs available at test time. However, when the extra training-only information is weak or noisy, this two-step approach can mislead the deployable model and even hurt accuracy. We propose a joint training method that learns the two models together, so the deployable model can benefit from the extra information only when it actually helps, instead of inheriting its mistakes. We provide guarantees that describe when joint training improves prediction accuracy and analyze a simple alternating training algorithm for large, high-dimensional models. Experiments on synthetic data and medical prediction tasks show that our approach avoids these failures and consistently outperforms standard two-step baselines.

This work proposes a novel method for solving learning from label proportion problems. For this purpose, we learn a classifier that minimizes three key objectives: (i) a bag-level loss, which quantifies the discrepancy between true and predicted label proportions in bags, (ii) an instance-level loss, inspired from domain adaptation, which leverages anchor samples with known labels and trainable supports and (iii) a distribution discrepancy that aims at aligning anchor's learned support with those of the bag samples. The problem is formulated as an alternating optimization process, iteratively updating the classifier and aligning distributions via a particle flow method. The flow of anchor samples is governed by a vector field designed to minimize the anchor loss while ensuring alignment between anchor and bag distributions. We provide a theoretical analysis, guaranteeing the convergence of the flow and identifying conditions under which the method achieves effective alignment. Our analysis highlights that gap and diversity in label proportions within bags is a critical factor for learnability. Empirical results on tabular and image datasets demonstrate the method's effectiveness, outperforming state-of-the-art approaches.

Low-Rank Adaptation (LoRA) has become the de facto paradigm for parameter-efficient fine-tuning, with its effectiveness critically influenced by rank allocation across modules. However, existing approaches face a fundamental dilemma: uniform allocation ignores module heterogeneity, while adaptive methods introduce expensive training overhead or lack reusability across configurations. We propose CSPLoRA (Confidence-guided Structural Planning for LoRA), a decoupled framework that reweights probe samples by prediction uncertainty to obtain more discriminative module importance estimates. The key insight is that hard samples---those the model struggles with---provide more informative gradient signals for identifying critical modules than easy samples. For a fixed task-model pair, the resulting structural priors can be reused across compatible rank budgets and LoRA backends, supporting a practical "probe once, deploy everywhere" workflow. Experiments on GLUE, commonsense reasoning, and arithmetic tasks show that CSPLoRA improves over uniform LoRA on average (+1.25 points on LLaMA-2-7B commonsense reasoning) while maintaining comparable parameters, with the planned rank structure reusable across compatible LoRA variants.

Simulating trajectories of dynamical systems is a fundamental problem in a wide range of fields such as molecular dynamics, biochemistry, and pedestrian dynamics. Machine learning has become an invaluable tool for scaling physics-based simulators and developing models directly from experimental data. In particular, recent advances in deep generative modeling and geometric deep learning enable probabilistic simulation by learning complex trajectory distributions while respecting intrinsic permutation and time-shift symmetries. However, trajectories of N-body systems are commonly characterized by high sensitivity to perturbations leading to bifurcations, as well as multi-scale temporal and spatial correlations. To address these challenges, we introduce STFlow (Spatio-Temporal Flow), a generative model based on graph neural networks and hierarchical convolutions. By incorporating data-dependent couplings within the Flow Matching framework, STFlow denoises starting from conditioned random-walks instead of Gaussian noise. This novel informed prior simplifies the learning task by reducing transport cost, increasing training and inference efficiency. We validate our approach on N-body systems, molecular dynamics, and human trajectory forecasting. Across these benchmarks, STFlow achieves the lowest prediction errors with fewer simulation steps and improved scalability.

Spatial reasoning from egocentric videos is inherently challenging because the observable evidence is constrained by the camera trajectory. Existing methods rely on single-turn inference, forcing models to resolve geometric ambiguity through semantic priors rather than verifiable evidence. We argue that spatial reasoning should be revisitable: conclusions formed under limited evidence should remain open to revision when complementary viewpoints become available. Building on this insight, we propose Reason, then Re-reason (ReRe), a training-free, inference-time framework with two phases: in the Reason Phase, an MLLM forms a spatial hypothesis from the original video; in the Re-reason Phase, it verifies or revises the hypothesis by observing a synthesized novel-view video. To enable effective cross-view revisiting, we design a Geometry-to-Video pipeline that renders strategically complementary novel views from predicted 3D geometry. These views feature an elevated, oblique perspective with scene-spanning coverage, while preserving the MLLM's native video interface without architectural modifications. Extensive evaluations on VSI-Bench and STI-Bench demonstrate that ReRe substantially boosts open-source MLLMs to rival proprietary state-of-the-art performance.

Discovering heterogeneous catalysts tailored for specific reaction intermediates remains a fundamental bottleneck in materials science. While traditional trial-and-error methods and recent generative models have shown promise, they struggle to capture the intrinsic coupling between surface geometry and adsorbate interactions. To address this limitation, we propose CatFlow, a flow matching-based framework for de novo design and structure prediction of heterogeneous catalysts. Our model operates on a primitive cell-based factorized representation of the slab-adsorbate complex, reducing the number of learnable variables by an average of 9.2x while explicitly encoding the surface orientation of the slab-adsorbate interface. Experiments on the Open Catalyst 2020 dataset demonstrate that CatFlow significantly improves the structural fidelity of generated catalysts compared to autoregressive and sequential baselines. Further experiments show that the generated structures accurately capture the adsorption energy distributions of physically plausible interfaces and lie closer to thermodynamic local minima.

Safety alignment of large language models is fragile: even small fine-tuning perturbations elastically revert behaviors toward those of the pre-training, with degradation inversely proportional to the size of the alignment set. We ask how to achieve safety alignment with minimal augmentation. To this end, we model augmentation as a set of group actions on sequences and formalize robustness gains as a normalized, monotone submodular function over transformations. We then leverage submodular optimization to select minimal augmentations that provably improve robustness. Experiments confirm that our approach efficiently restores safety alignment while minimizing the overhead of augmentation.

Accurate survival modeling in biomedical studies is often hindered by rare events, limited effective sample sizes, and settings with limited or partially observed information (e.g., covariates of interest that are difficult or expensive to collect, highly structured sampling designs, or nuisance parameters omitted by conditioning). Knowledge distillation can leverage external predictive information without sharing individual-level data, but existing approaches are largely built for fully specified likelihoods or probability-based survival models and do not extend to settings where outcome distributions are only partially specified. To address this challenge, we propose a knowledge distillation framework based on a composite-likelihood Kullback--Leibler divergence that aligns teacher and student models within components. Our key insight is that, although composite likelihoods do not define a global outcome distribution, each likelihood component induces a well-defined probability model on its restricted outcome space, enabling a principled KL divergence. Simulation studies and biomedical case studies show improved discrimination and predictive accuracy in rare-event, heterogeneous settings without requiring access to external individual-level data.

Offline-to-online reinforcement learning (O2O RL) faces a central challenge between retaining offline conservatism and adapting to online feedback under distribution shift. This challenge arises because data behavior evolves during fine-tuning, rendering data origin a misleading basis for constraint handling and thereby leading to objective–data mismatch. We therefore propose Dynamic Alignment for RElaxation (DARE), a distribution-aware framework for sample-level constraint relaxation based on the behavioral consistency with a behavior model. To our knowledge, DARE is the first to condition constraint relaxation on behavioral consistency via a posterior-induced exchange mechanism, moving beyond a binary offline/online data distinction. Importantly, DARE requires only per-sample behavioral alignment, enabling instantiation on top of many offline algorithms with flexible choices of behavior models and fine-tuning objectives. We provide a theoretical analysis showing that behavior-based sample exchange consistently improves the distinction between offline-like and online-like subsets. Experiments on D4RL demonstrate that DARE consistently improves fine-tuning stability and achieves superior final performance over strong offline-to-online baselines. (The code is publicly available at \url{https://github.com/lpzu/DARE}.)

We reconceptualize Personalized Multimodal Federated Learning (PMFL) by treating missing modalities as intrinsic structural identities that constrain each client to a distinct Riemannian submanifold, rather than as deficiencies to be compensated. To reconcile the tension between identity preservation and cross-client collaboration, we cast PMFL as an identity-aware potential game and seek a geometry-consistent equilibrium instead of a monolithic full-modality optimum. We propose GeoEvo, a federated approximate solver that combines curvature-adaptive Fisher descent with manifold-lifted evolutionary search: Natural Evolution Strategies for basin escape, and Particle Swarm updates anchored to a server-broadcast Fr\'echet prototype for cross-client transfer. A monotone acceptance rule drives per-step potential dissipation, yielding an $\mathcal{O}(1/\sqrt{T})$ stationarity rate and convergence toward first-order Nash equilibria in non-convex regimes. Empirically, GeoEvo improves personalization and robustness across diverse modality-missing patterns.

Pairwise preference learning is central to machine learning, with recent applications in aligning language models with human preferences. A typical dataset consists of triplets $(x, y^+, y^-)$, where response $y^+$ is preferred over response $y^-$ for context $x$. The Bradley--Terry (BT) model is the predominant approach, modeling preference probabilities as a function of latent score differences. Standard practice assumes data follows this model and learns the latent scores accordingly. However, real data may violate this assumption, and it remains unclear what BT learning recovers in such cases. Starting from triplet comparison data, we formalize the preference information it encodes through the conditional preference distribution (CPRD). We give precise conditions for when BT is appropriate for modeling the CPRD, and identify factors governing sample efficiency---namely, margin and connectivity. Together, these results offer a data-centric foundation for understanding what preference learning actually recovers.