ICML 2026 Papers

The field of computer architecture, which bridges high-level software abstractions and low-level hardware implementations, remains absent from current large language model (LLM) evaluations. To this end, we present QuArch (pronounced ‘quark’), the first benchmark designed to facilitate the development and evaluation of LLM knowledge and reasoning capabilities specifically in computer architecture. QuArch v1.0 provides a comprehensive collection of 2,671 expert-validated question-answer (QA) pairs covering various aspects of computer architecture, including processor design, memory systems, and interconnection networks. Our evaluation reveals that while frontier models possess domain-specific knowledge, they struggle with skills that require higher-order thinking in computer architecture. Frontier model accuracies vary widely (from 34% to 73%) on these advanced questions, highlighting persistent gaps in architectural reasoning across analysis, design, and implementation QAs. Furthermore, via fine-tuning we find that QuArch can translate to improved performance on a realistic memory hierarchy design task, resulting in up to 1.99× more area-efficient solutions and up to 40% more viable solutions overall. By holistically assessing fundamental skills, QuArch provides a foundation for building and measuring LLM capabilities that can accelerate innovation in computing systems. The QuArch benchmark and leaderboard are publicly available at: https://quarch.ai/.

Generative models are increasingly paired with safety classifiers that filter harmful or undesirable outputs. A common strategy is to fine-tune the generator to reduce the probability of being filtered, but this can be suboptimal: it often pushes the model toward producing samples near the classifier’s decision boundary, increasing both false positives and false negatives. We propose Boundary Guidance, a reinforcement learning fine-tuning method that explicitly steers generation away from the classifier’s margin. On a benchmark of jailbreak, ambiguous, and long-context prompts, Boundary Guidance improves the safety while maintaining or improving the utility of outputs, as judged by LLM-as-a-Judge evaluations. Comprehensive ablations across model scales and reward designs demonstrate the robustness of our approach.

We study the role of batch size in stochastic conditional gradient methods under a $\mu$-Kurdyka–Łojasiewicz ($\mu$-KL) condition. Focusing on momentum-based stochastic Frank–Wolfe–type conditional gradient algorithms (e.g., Scion), we derive a new analysis that explicitly captures the interaction between stepsize, batch size, and stochastic noise. Our study reveals a regime-dependent behavior: increasing the batch size initially improves optimization accuracy, but beyond a critical threshold, the benefits saturate and can eventually degrade performance under a fixed token budget. Notably, the theory predicts the magnitude of the optimal stepsize and aligns well with empirical practices observed in large-scale training. Leveraging these insights, we derive principled guidelines for selecting the batch size and stepsize, and propose an adaptive strategy that increases batch size and sequence length during training while preserving convergence guarantees. Preliminary experiments are consistent with the theoretical predictions and illustrate the emergence of the predicted scaling regimes. Overall, our results provide a theoretical framework for understanding batch-size scaling in stochastic conditional gradient methods and offer guidance for designing efficient training schedules in large-scale optimization.

Closed-loop evaluation of autonomous-driving policies requires interactive simulation beyond log replay. Existing generative world models suffer three gaps: history-incompatible initialization, sampling latency exceeding real-time budgets, and compounding kinematic infeasibility. We propose VectorWorld, a streaming vector-graph world model that incrementally generates ego-centric lane--agent tiles during rollout. VectorWorld couples a motion-aware gated VAE for history-compatible initialization, an edge-gated relational DiT with interval-conditioned MeanFlow and JVP-based large-step supervision for solver-free outpainting, and $\Delta$Sim, a physics-aligned NPC policy with hybrid discrete--continuous actions and differentiable kinematic logit shaping. On Waymo Open Motion and nuPlan, VectorWorld improves map fidelity, initialization validity, and density calibration, enabling stable real-time $1\mathrm{km}+$ closed-loop rollouts.

Deep learning-based methods have shown remarkable effectiveness in solving PDEs, largely due to their ability to enable fast simulations once trained. However, despite the availability of high-performance computing infrastructure, many critical applications remain constrained by the substantial computational costs associated with generating large-scale, high-quality datasets and training models. In this work, inspired by studies on the structure of Green's functions for elliptic PDEs, we introduce Neural-HSS, a parameter-efficient architecture built upon the Hierarchical Semi-Separable (HSS) matrix structure that is provably data-efficient for a broad class of PDEs. We theoretically analyze the proposed architecture, proving that it satisfies exactness properties even in very low-data regimes. We also investigate its connections with other architectural primitives, such as the Fourier neural operator layer and convolutional layers. We experimentally validate the data efficiency of Neural-HSS on the three-dimensional Poisson equation over a grid of two million points, demonstrating its superior ability to learn from data generated by elliptic PDEs in the low-data regime while outperforming baseline methods. Finally, we demonstrate its capability to learn from data arising from a broad class of PDEs in diverse domains, including electromagnetism, fluid dynamics, and biology.

Graph neural networks are widely used for node classification, but they remain vulnerable to out-of-distribution (OOD) shifts in node features and graph structure. Prior work established that methods trained with standard supervised learning (SL) objectives tend to capture spurious signals from either features and/or structure, leaving the model fragile under distributional changes. To address this, we propose TIDE, a novel and effective Tri-Component Information Decomposition framework that explicitly decomposes information into feature-specific, structure-specific and joint components. TIDE aims to preserve only the label-relevant part of the joint information while filtering out spurious feature- and structure-specific information, thereby enhancing the separation between in-distribution (ID) and OOD nodes. Beyond the framework, we provide theoretical and empirical analyses showing that an information bottleneck objective is preferable to standard SL for graph OOD detection, with higher ID confidence and a greater entropy gap between ID and OOD data. Extensive experiments across seven datasets confirm the efficacy of Tide, achieving up to a 34% improvement in FPR95 over strong baselines while maintaining competitive ID accuracy. Code is available at https://github.com/DannyW618/TIDE.

Asynchronous reinforcement learning (RL) has shown notable success in accelerating the post-training of large language models (LLMs). However, its decoupled data generation and training paradigm introduces a fundamental distributional mismatch between data generated by stale behavior policies and current policy, leading to unstable training and degraded performance. To address this challenge, we propose D-ARL, a **D**istribution-matched **A**synchronous **R**einforcement **L**earning framework that selects high-quality asynchronous samples whose distributions are well aligned with the current policy for policy optimization. Specifically, D-ARL maintains a replay buffer that collects samples from the most recent $K$ behavior policies and proposes a variance-guided metric to select distribution-matched data. During training, D-ARL introduces a multi-behavior policy optimization algorithm to leverage the multi-source nature of the selected samples for policy update. Experiments on six widely used reasoning benchmarks show that D-ARL outperforms state-of-the-art asynchronous methods, achieving an average improvement of 6.4\% in reasoning performance and 34.7\% in sample efficiency.

Learning latent dynamics is central to assessing current states and forecasting future trajectories for high-dimensional time series. For locally-stationary latent dynamics parameterized by past latent states and an environment variable c, with latent dynamics state z_t, prior identifiability results largely focus on z_t when conditioned on pre-defined label u of the dynamics environment. This leaves two limitations: reliance on pre-defined labels that hinder generalization to unseen environments, and limited understanding of the identifiability of F and c which---while offering important structural properties for the identifiability of z_t---are learned jointly with z_t. We address these challenges with Meta-iLaD, a novel latent dynamics framework to attain identifiability by meta-learning across dynamics environments. Meta-iLaD replaces the conditioning of c on pre-defined labels with a novel conditional prior, modeled as a feedforward meta-learner that rapidly extracts c from few-shot examples. We further establish simultaneous identifiability for z_t, c and F, for a general formulation of the dynamics function F over past latent states and c, without restricting the dimension of c or how it modulates F. We provide strong empirical evidence that 1) conditioning on few-shot examples enables generalization to out-of-distribution environments, and 2) identifiability for c and F is critical for accurate forecasting beyond reconstructing observed trajectories.

Vision-Language Pre-trained (VLP) models are vulnerable to real-world query noise. Current cross-modal Test-Time Adaptation (TTA) methods often rely on high-confidence predictions, which induces confirmation bias and neglects the informative signals in ambiguous Low-Confidence Queries. To address this, we propose Generative Semantic Refinement and Exclusion-Guided Adaptation (ReEx), a robust retrieval framework that extends adaptation to the entire query stream. Specifically, textual structural noise is rectified by a Generative Semantic Refinement (GSR) module, which employs Confidence-Guided Dynamic Fusion to anchor LLM-based repairs and prevent semantic drift. To exploit ambiguous data, adaptation is driven by Exclusion-Guided Proxy Contrastive Learning (EPCL), which imposes negative constraints via Exclusion Sets of unlikely candidates. Experimental results on COCO-C and Flickr-C demonstrate that ReEx consistently outperforms existing TTA methods, achieving significant robustness gains with a justifiable computational trade-off.

Causal discovery is a difficult problem that typically relies on strong assumptions on the data-generating model, such as non-Gaussianity. In practice, many modern applications provide multiple related views of the same system, which has rarely been considered for causal discovery. Here, we leverage this multi-view structure to achieve causal discovery with weak assumptions. We propose a multi-view linear Structural Equation Model (SEM) that extends the well-known framework of non-Gaussian disturbances by alternatively leveraging correlation over views. We prove the identifiability of the model for acyclic SEMs. Subsequently, we propose several multi-view causal discovery algorithms, inspired by single-view algorithms (DirectLiNGAM, PairwiseLiNGAM, and ICA-LiNGAM). The new methods are validated through simulations and applications on neuroimaging data, where they enable the estimation of causal graphs between brain regions.

Designing heuristic algorithms for complex optimization problems is a time-consuming and expert-driven process. Recently, Automated Heuristic Design (AHD) using Large Language Models (LLMs) has shown significant promise for automating algorithm development. However, existing works mainly rely on programs to represent heuristics, which are inherently taskspecific and fail to generalize as effectively as established metaheuristics like tabu search or guided local search. To bridge this gap, we introduce Multi-Task Hierarchical Search (MTHS), an LLM-guided evolutionary method that co-designs general-purpose metaheuristics and task-specific programs. MTHS employs a hierarchical representation and adopts a two-level evolution framework to evolve task-agnostic metaheuristics and task-specific program implementations simultaneously across multiple heuristic design tasks. During this evolution, a knowledge transfer mechanism allows learning from elite programs designed for other tasks. We evaluated MTHS on distinct combinatorial optimization problems, where it outperforms both commonly-used heuristics and existing LLM-driven AHD approaches. Our results demonstrate that the hierarchical representations facilitate effective multi-task AHD, and the evolved metaheuristics exhibit strong generalization to related tasks. Code: https://github.com/FeiLiu36/MTHS.

Data heterogeneity is a fundamental challenge in Federated Learning (FL), where induced model drift often results in "negative gains" for global models on data-abundant clients, with performance falling below that of local training. To address this issue, we propose FedGain, a novel framework that optimizes collaborative client clustering to mitigate the negative gain. We are the first to develop a modified Scaling Law (SL) to quantify the reduction in data utility caused by heterogeneity and define Effective Federated Capacity to align clients with the highest potential collaboration gains. Extensive experiments demonstrate that our modified SL strictly adheres to the power-law learning discipline in non-IID scenarios. FedGain effectively suppresses negative gains to a negligible level across various FL algorithms and outperforms other Clustered FL methods.

While self-consistency methods have emerged as a promising approach to enhance the correctness of large language model (LLM) outputs by aggregating multiple stochastic samples, they suffer from two critical limitations, resulting in high computation cost. First, they evaluate output consistency monolithically, failing to efficiently combine partially correct answers across multiple samples. Second, they use static stopping criteria that cannot adapt to varying task complexities and model capabilities, resulting in suboptimal computational efficiency.In this work, we present Task-and-Model-Aware Fractal-Consistency (TMAFC), a novel self-consistency framework that addresses these limitations through two key innovations: (1) Fractal-Consistency, which evaluates output consistency at the granularity of output components to effectively combine partial correct answers across samples, and (2) Adaptive Stopping Criteria Calibration (ASCC), which dynamically adjusts sampling stopping criteria based on real-time assessment of both task difficulty and LLM capability. Through extensive experiments on diverse question-answering benchmarks, we demonstrate that TMAFC achieves superior efficiency-accuracy trade-offs, reducing sample cost by up to 55% while maintaining competitive accuracy compared to state-of-the-art baselines.

Continuous Normalizing Flows (CNFs) enable elegant generative modeling but remain bottlenecked by their iterative nature requiring costly sampling and lacking interpretability of the intermediate states. Recent approaches accelerate sampling by straightening trajectories or distilling endpoints, yet they treat the original generative process as a black box, discarding the teacher’s intermediate dynamics. We propose a fundamentally different perspective: globally linearizing flow dynamics via Koopman theory to achieve trajectory-preserving linearization. By lifting Conditional Flow Matching (CFM) into a higher-dimensional Koopman space, we represent its evolution with a single linear operator. Crucially, unlike boundary-only distillation, our method enforces infinitesimal consistency with the teacher's vector field along *the full generative path*. We derive a practical, simulation-free training objective that ensures this global alignment and yields two key benefits. First, sampling becomes one-step and parallelizable. Second, because the linearization is faithful to the dynamics, the Koopman operator provides unique insights on the generation. We demonstrate that this structure enables novel applications unavailable in prior approaches, including discovery of semantically coherent editing directions, inversion with a teacher-aligned linear operator and class-conditional spectral signatures. Empirically, our approach achieves competitive sample quality, while enabling spectral analysis and control of the *entire trajectories* of generative flows.

Metaphorical videos are prevalent across various real-world scenarios to convey complex ideas, and understanding them typically requires high-order cognitive capabilities. The lack of systematic studies on metaphorical video understanding not only constrains the real-world applicability of MLLMs but also impedes the thorough assessment of their high-order cognitive capabilities. To bridge this gap, we propose MetaphorVU-Bench, the first systematic and comprehensive benchmark dedicated to metaphorical video understanding. Through experiments, we find current MLLMs struggle with accurate metaphorical video understanding, lagging far behind human level, primarily due to defective cross-domain mapping. Motivated by this finding, we construct a metaphor knowledge graph as mapping augmentation and propose MetaphorBoost, an inference-time enhancement framework achieving consistent performance improvement. Our benchmark, analysis, and method provide useful insights and a foundation for future research on advancing MLLMs. Code: https://github.com/icip-cas/MetaphorVU.

Vision-Language-Action (VLA) models show strong generalization for robotic control, but finetuning them with reinforcement learning (RL) is constrained by the high cost and safety risks of real-world interaction. Training VLA models in interactive world models avoids these issues but introduces several challenges, including pixel-level world modeling, multi-view consistency, and compounding errors under sparse rewards. Building on recent advances across large multimodal models and model-based RL, we propose VLA-MBPO, a practical framework to tackle these problems in VLA finetuning. Our approach has three key design choices: (i) adapting unified multimodal models (UMMs) for data-efficient world modeling; (ii) an interleaved view decoding mechanism to enforce multi-view consistency; and (iii) chunk-level branched rollout to mitigate error compounding. Theoretical analysis and experiments across simulation and real-world tasks demonstrate that VLA-MBPO significantly improves policy performance and sample efficiency. Crucially, our method maintains a universal set of hyperparameters across all tasks, underscoring its robustness and scalability for real-world robotic deployment.

Low-rank approximation has emerged as a cornerstone technique for model compression and parameter-efficient fine-tuning, enabling substantial reductions in computation and memory without altering model architectures. However, existing approaches often overlook the shifts in feature distributions induced by the approximation process, which can lead to error amplification and unstable inference. We propose a distribution-aware whitening framework that dynamically whitens layer inputs based on the evolving feature distributions, ensuring second-order isotropy of input features. This allows that the discarded components in the low-rank approximation are those with minimal impact on model outputs, thereby minimizing cumulative approximation errors across layers. We theoretically analyze how distribution misalignment leads to error propagation and demonstrate that our approach achieves tighter control over layerwise distortion. Extensive experiments across various large language models demonstrate the superiority of our method in post-training compression. Moreover, our method can also serve as an effective initialization for LoRA-style parameter-efficient fine-tuning. Our findings highlight the importance of considering feature distributions in low-rank approximations, paving the way for reliable and effective model compression strategies.

AI developers face a dual-use dilemma. A model capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limiting dual-use AI capabilities to trusted deployments with a legitimate need. A gold standard for access control would be to serve separate models with different capabilities to different users. However, training and deploying multiple models is prohibitively expensive. To address this challenge, we propose gradient-routed auxiliary modules (GRAM), a pre-training method that adds modules to a neural network and selectively updates them to induce specialization. Ablating a module at inference time removes its capability from the network, approximating a model trained on filtered data. We evaluate GRAM on synthetic stories and realistic dual-use data spanning virology, cybersecurity, nuclear physics, and specialized code. These experiments show that GRAM preserves selected retain capabilities while disabling forgotten capabilities, and limits recovery better than post-hoc unlearning. Most importantly, a Chinchilla-optimal scaling analysis from 50M to 5B parameters shows that the forget capability gap between data-filtered and full-data models widens with scale while the retain gap stays constant, and that GRAM closely tracks data filtering. GRAM's training cost is independent of the number of supported capability profiles, yielding a 5x cost reduction in our 5-profile setting.

Recent advances in audio deepfake detection have been driven by increasingly large speech foundation models and growing amounts of synthetic data. Despite strong benchmark performance, it remains unclear how detection capability scales with model capacity and training data under realistic deployment conditions involving distribution shift, signal corruption, and unseen synthesis pipelines. In this work, we present the first systematic study of scaling laws in post-training audio deepfake detection, focusing on fine-tuning regimes rather than large-scale pretraining. Using a controlled family of speech foundation models with shared architecture and pretraining, we analyze how detection performance, robustness, and generalization evolve as a function of model size and training data scale. Our results reveal a fundamental asymmetry between performance scaling and robustness scaling in audio deepfake detection, suggesting increasing model capacity alone is insufficient for achieving reliable real-world generalization.