ICML 2026 Papers

Aligning large language models (LLMs) to heterogeneous and rapidly evolving safety requirements remains a critical challenge. Existing instruction-tuned LLMs and standalone safety classifiers often fail to generalize to new safety configurations, motivating the need for Reward Models (RMs) that are explicitly configurable to changing specifications. We introduce the Configurable Safety Reward Model (CSRM), which is jointly optimized for calibrated safety compliance and reward modeling. Our approach is supported by configuration-targeted data augmentation that enforces instruction adherence while preserving relative severity structure. The resulting RM is sensitive to fine-grained safety configurations and conversational nuances, substantially improving generalization to previously unseen safety configurations. CSRM achieves state-of-the-art performance on recent configurable safety benchmarks, including CoSApien (94.6\% F1) and DynaBench (75.8\% F1), without requiring additional human annotation. When used for downstream safety alignment, CSRM yields LLMs with a significantly improved helpfulness–safety tradeoff compared to existing baselines.

Continual learning with large pre-trained models offers significant potential for cross-task knowledge accumulation, but faces critical challenges such as catastrophic forgetting and parameter interference, especially when historical data is unavailable. Existing approaches typically rely on sequential fine-tuning or model merging strategies, yet often overlook the impact of loss landscape sharpness and dominant singular value directions, which leads to subspace misalignment and severe knowledge forgetting. In this paper, we propose the Sharpness-Aware Isotropic Merging (SAIM) framework, which introduces targeted optimizations in both the fine-tuning and merging stages to address these issues. Specifically, SAIM consists of two synergistic modules: (1) a Sharpness-Aware Block Coordinate Descent (SA-BCD) optimizer that guides the model toward flatter minima and selectively updates the most task-sensitive parameters, thereby mitigating parameter interference and enhancing robustness; (2) an adaptive isotropic merging algorithm that dynamically balances the singular value spectrum across tasks, effectively preventing the model from overemphasizing any single task direction, maintaining balanced knowledge representation, and improving subspace alignment. Extensive experiments on vision and language benchmarks demonstrate that SAIM achieves 5-10\% higher accuracy than existing methods and maintains robust performance as the number of tasks increases. Our code are available at \url{https://github.com/Yangqun123456/SAIM}.

Real-world perception and decision making are inherently multimodal, integrating complementary signals across modalities. However, training multimodal models faces two main obstacles. First, collecting large-scale, well-aligned paired multimodal datasets is often impractical, making end-to-end multimodal training difficult. Second, existing multimodal representations frequently entangle information shared across modalities with modality-specific information, hindering interpretability and control. We introduce MultiLoReFT, an efficient and scalable low-rank representation fine-tuning framework for multimodal learning with pretrained unimodal models. MultiLoReFT extends low-rank adaptation to the multimodal setting and learns interpretable projection subspaces that decouple shared and modality-specific information. Across simulated and real-world benchmarks, it produces representations that support multimodal prediction while explicitly revealing how shared and modality-specific information is distributed across modalities.

Existing online change-point detection (CPD) methods rely on fixed-dimensional Euclidean summaries, implicitly assuming that distributional changes are well captured by moment-based or feature-based representations. They can obscure important changes in distributional shape or geometry. We propose an intrinsic distribution-valued CPD framework that treats streaming batch data as a stochastic process on the 2-Wasserstein space. Our method detects changes in the law of this process by mapping each empirical distribution to a tangent space relative to a pre-change Fréchet barycenter, yielding a reference-centered local linearization of 2-Wasserstein space. This representation enables sequential detectors by adapting classical multivariate monitoring statistics to tangent fields. We provide theoretical guarantees and demonstrate, via synthetic and real-world experiments, that our approach detects complex distributional shifts with reduced detection delay at matched $\mathrm{ARL}_0$ compared with moments-based and model-free baselines. The code is available at https://github.com/yyzeng43/IDD-icml.

Generative Flow Networks (GFlowNets) have emerged as a powerful framework for molecular generation, sampling diverse candidates proportionally to a reward function. However, the vast chemical space necessitates truncating trajectory length, forcing models to treat incomplete molecular fragments as terminal states alongside valid molecules. This conflation distorts the learned distribution by allocating probability mass to chemically meaningless states. We propose LeakGFN, a dual-head architecture that decomposes flow into two components: a chemical head modeling flow over the full chemical space, and a valid head estimating the fraction of flow reaching valid molecules within the truncation boundary. Through this decomposition, the valid head implicitly learns molecular reachability without explicit supervision. We prove that LeakGFN recovers the correct distribution over accessible molecules under mild assumptions. Experiments on five molecular optimization tasks demonstrate consistent improvements over flow matching baselines, achieving state-of-the-art performance on four out of five tasks. Our module integrates as a plug-and-play enhancement into existing frameworks, improving performance on both pocket-conditioned and multi-objective generation tasks.

Spiking Neural Networks (SNNs) enable event-driven computation with sparse activations, but building multimodal Transformers on SNNs is hindered by unstable training in deep spiking stacks and a mismatch between dense softmax attention and spike-based communication. We propose SMM Transformer, an SNN-based multimodal Transformer framework that combines (i) a Parallel LIF with Multistage Learnable Parameters (PLMP) neuron and a tailored P-STBP algorithm to stabilize training, (ii) a spike-driven attention approximation (SMSA) with a lightweight self-compensation branch, and (iii) a spiking mixture-of-experts (SMoE) module for modality-aware fusion. Across visual and multimodal benchmarks, SMM Transformer achieves competitive accuracy compared to ANN baselines while reducing the estimated compute energy of the attention module by up to 97\% under a standard MAC/AC cost model.

Prompt learning for vision-language models (VLMs) primarily follows end-to-end or decoupled routes to balance base and new task performance, but suffers a fundamental bottleneck: sample-wise optimization within task-specific feature spaces traps models in local optima, hindering global optimality. To address this, we identify a key insight that VLMs can be prompted within a Coupled Prompt Field-a shared space where base and new tasks are mutually constrained-and present AlignedNorm, which enforces the field coupling. By dynamically aligning the norms of prompts to VLMs' native scale, our method enables joint optimization of both tasks. Without complex designs, our method matches leading decoupled approaches on 15 datasets across 4 experimental settings, offering both a new perspective and a practical solution to the local-optima dilemma in prompt learning.

Omni-modal Large Language Models (Omni-LLMs) have demonstrated strong capabilities in audio-video understanding tasks. However, their reliance on long multimodal token sequences leads to substantial computational overhead. Despite this challenge, token compression methods designed for Omni-LLMs remain limited. To bridge this gap, we propose OmniSIFT (Omni-modal Spatio-temporal Informed Fine-grained Token compression), a modality-asymmetric token compression framework tailored for Omni-LLMs. Specifically, OmniSIFT adopts a two-stage compression strategy: (i) a spatio-temporal video pruning module that removes video redundancy arising from both intra-frame structure and inter-frame overlap, and (ii) a vision-guided audio selection module that filters audio tokens. The entire framework is optimized end-to-end via a differentiable straight-through estimator. Extensive experiments on five representative benchmarks verify the efficacy and robustness of OmniSIFT. Notably, for Qwen2.5-Omni-7B, OmniSIFT adds 4.85M parameters while still achieving lower latency than training-free baselines such as OmniZip. With only 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the full-token model on several tasks.

Effectively scaling Reinforcement Learning (RL) is crucial for enhancing the reasoning and alignment of Large Language Models. The massive data and complex execution flows inherent in these tasks require a distributed architecture capable of efficient scaling. However, to simplify programming and dependency management, mainstream frameworks often rely on a centralized architecture where a single node dispatches both control and data. This inherent coupling creates significant communication bottlenecks, severely limiting system scalability and efficiency. We present DistFlow, a novel, fully distributed RL framework that adopts a multi-controller paradigm. By decoupling data transmission from control dispatch, DistFlow establishes a parallelism-aware, decentralized Data Coordinator that leverages local caching, load balancing, and asynchronous double buffer to minimize communication overhead and mitigate straggler effects. For control logic, it introduces a task scheduler built upon Directed Acyclic Graph (DAG) that facilitates fine-grained, independent execution. Experimental results demonstrate that DistFlow achieves near-linear scalability up to 512 GPUs and delivers up to a 2.63x throughput improvement over state-of-the-art (SOTA) frameworks.

We present IQA-Spider, the first image quality assessment (IQA) framework that unifies reasoning, grounding, and referring into a single LMM-based framework for multi-granularity quality understanding. Existing LMM-based IQA methods typically support only partial perception dimensions, such as quality description and question answering~(\textit{i.e.}, reasoning) or pixel-level grounding. This limitation largely stems from the absence of (i) a unified task and data formulation and (ii) effective optimization paradigms for multi-granularity learning. To address these limitations, we formulate a rigorous four-task paradigm covering global and local quality description, pixel-level grounding, and region-level referring. Based on this formulation, we construct a corresponding IQA dataset with a scalable and automatic annotation pipeline, thereby providing a solid foundation for unified multi-granularity learning. To further enable unified perception, we adopt a conflict-free two-stage design that progressively extends text-level multi-granularity understanding to pixel-level grounding: (i) the first stage equips the model with fine-grained text-level reasoning across multiple IQA tasks, and (ii) the second stage introduces a training-free text-to-point grounding paradigm, which bridges textual semantics and pixel-level perception by mapping token logits to spatial coordinates. Based on these efforts, we achieve IQA-Spider with unified multi-granularity explainable image quality assessment. Extensive experiments across multiple benchmarks demonstrate strong performance, validating the effectiveness and versatility of the proposed formulation and framework. Our code and dataset will be released at: \href{https://github.com/Helen1p/IQA-Spider.git}{https://github.com/Helen1p/IQA-Spider.git}.

It is infeasible to encompass all possible disturbances within the training dataset. This raises a critical question regarding the robustness of Vision-Language-Action (VLA) models when encountering unseen real-world visual disturbances, particularly under imperfect visual conditions. In this work, we conduct a systematic study based on recent state-of-the-art VLA models and reveal a significant performance drop when visual disturbances absent from the training data are introduced. To mitigate this issue, we propose a lightweight adapter module grounded in information theory, termed the Information Bottleneck Adapter (IB-Adapter), which selectively filters potential noise from visual inputs. Without requiring any extra data or augmentation strategies, IB-Adapter consistently improves over the baseline by an average of 30%, while adding fewer than 10M parameters, demonstrating notable efficiency and effectiveness. Furthermore, even with a 14x smaller backbone (0.5B parameters) and no pre-training on the Open X-Embodiment dataset, our model StableVLA achieves robustness competitive with 7B-scale state-of-the-art VLAs. With negligible parameter overhead (<10M), our approach maintains accuracy on long-horizon tasks and surpasses OpenPi under real-world visual disturbances. The code will be made publicly available.

As Large Language Models (LLMs) saturate elementary benchmarks, the research frontier has shifted from generation to the reliability of automated evaluation. We demonstrate that standard "LLM-as-a-Judge" protocols suffer from a systematic evaluation Alignment Gap when applied to upper-undergraduate to early graduate level mathematics. To quantify this, we introduce QEDBench, the first benchmark to systematically measure alignment with human experts on undergraduate-level math proofs by contrasting course-specific rubrics against expert common knowledge criteria. By deploying a dual-evaluation matrix ($7$ judges $\times$ $5$ solvers) against 1,000+ hours of human evaluation, we reveal that certain frontier evaluators like Claude 4.5 Opus exhibit significant positive bias (up to $+0.28$ mean score inflation), effectively "hallucinating rigor" in flawed proofs. Furthermore, we uncover a critical reasoning disparity: while Gemini 3.0 Pro achieves state-of-the-art performance (0.91 raw score), specialized reasoning models like o3-deep-research collapse in discrete domains, dropping to 42.1\% accuracy in Graph Theory. We release QEDBench as a public benchmark for evaluating and improving AI judges.

Agentic Web is an emerging paradigm where autonomous agents help users use online information. As the paradigm develops, content providers are also deploying agents to manage their data and serve it through controlled interfaces. This shift moves information access from centralized retrieval to decentralized coordination. To study this setting, we introduce AgentWebBench, a benchmark that evaluates how well a user agent synthesizes answers by interacting with website-specific content agents. We evaluate four tasks that cover common web information needs, spanning ranked retrieval (web search, web recommendation) and open-ended synthesis (question answering, deep research). Across seven advanced LLMs and three coordination strategies, multi-agent coordination generally lags behind centralized retrieval as expected, because user agent cannot directly access the corpus, but the gap shrinks with model scale and can even outperform centralized retrieval on question answering. This benchmark also enables us to study properties of the emerging paradigm of the digital world. We find that decentralized access concentrates traffic toward a small set of websites, test time scaling improves both interaction reliability and task performance, and strong results require sufficient interactions guided by careful planning. Finally, our failure analysis suggests that user agents need better planning and answer synthesis, while content agents need more reliable retrieval and evidence quality. Code, data, and APIs are released on https://github.com/cxcscmu/AgentWebBench.

Sufficient dimension reduction (SDR) makes high-dimensional regression tractable by projecting the covariates onto a low-dimensional subspace that preserves the conditional mean of the response. Existing gradient-based estimators either operate in the ambient space and suffer from the curse of dimensionality, or localize in the reduced space at a per-outer-iteration cost at least quadratic in the sample size. We show that minimizers of the population Minimum Average Variance Estimation (MAVE) risk approximate the same Grassmannian target as the Outer Product of Gradients (OPG), and recast the empirical criterion as a smooth maximization on the Stiefel manifold with closed-form Riemannian gradient. The resulting algorithm, SMAVE, combines sparse projected-space nearest-neighbor localization with Riemannian stochastic gradient ascent. A simplified version comes with almost-sure convergence and a non-asymptotic rate matching the standard non-convex stochastic first-order scaling. Empirically, SMAVE matches or improves on RMAVE's synthetic subspace recovery at moderate-to-high ambient dimension, and on four real datasets it uniformly improves over OPG and is competitive with or outperforms RMAVE at orders of magnitude lower runtime.

Recent success of large reasoning models (such as OpenAI o1 and DeepSeek R1) have spurred a resurgence of interest in reinforcement learning from verifiable rewards (RLVR). However, progress is still largely driven by RL algorithm design, while data scheduling -- the data-side decisions that determine what the model trains on over time -- is critical but remains underexplored. Therefore, data scheduling becomes the focus of this paper, including how to curate data for supervised fine-tuning (SFT) and how to select prompts and collect rollouts for reinforcement learning (RL). We introduce a pipeline with careful designs on data scheduling, consisting of hardness-prioritized SFT and two-stage RL. Specifically, we first fine-tune the base model on supervision data that is curated to prioritize difficulty based on both arena learning and classification. Then, we introduce two-stage RL where a decreased max sequence length during rollout is used in the first stage to expand entropy and reduce repetition, and a large number rollouts per prompt and curriculum design are adopted in the second stage to encourage exploration for challenging problems. We implement this pipeline on Qwen2.5-32B and an internal 389B MoE model, and evaluate them on a wide range of benchmarks including challenging LeetCode and Codeforces weekly contests. The results not only indicate the effectiveness and scalability of our pipeline but also demonstrate our model achieve sota of 32B models in competitive code generation.

Optimizing OS kernels for specific applications is vital for peak performance, yet existing LLM-based methods struggle with a semantic mismatch between generalized reasoning and low-level system behaviors. As a result, these static, open-loop approaches suffer from runtime blindness, configuration fragmentation, and search drift, ultimately failing to unlock the system’s performance potential. To address this, we propose QiMeng-PerceptOS, an autonomous framework that shifts the paradigm to perception-driven tuning. QiMeng-PerceptOS integrates: (1) a Perception Module that aligns raw telemetry into high-fidelity semantic fingerprints; (2) a Global Search Module utilizing a Bi-level Hierarchical Induction Tree (BHIT) for global navigation and efficient pruning; and (3) a Posterior Enhancement Module to suppress hallucinations via trajectory synthesis. Experiments across diverse workloads show that it achieves significant performance breakthroughs by optimizing kernel configurations, reaching 296.6% of default Redis throughput and surpassing SOTA baselines by 32.6% within only 15 iterations. By establishing a perception-driven closed-loop, QiMeng-PerceptOS provides new insights for fully automated, large-scale system optimization.

Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation, featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and universal breakdown in musical pitch control.

Predicting high-dimensional transcriptional responses to genetic perturbations is challenging because signals are sparse and experimental noise is severe. Existing methods often suffer from mean collapse, achieving high correlation by predicting the global average expression rather than perturbation-specific responses, which yields false positives and poor interpretability. Methods that add biological knowledge graphs typically treat them as dense, static priors shared across perturbations, propagating noise. We propose AdaPert, which counters mean collapse by extracting a sparse, perturbation-specific subgraph via differentiable node selection, then suppressing spurious variation in non-responsive genes while emphasizing differentially expressed ones. Across multiple benchmarks, AdaPert outperforms existing baselines, with the largest gains on DEG-aware metrics.

Transformer attention is typically implemented using softmax normalization, which enforces attention weights with unit sum normalization. While effective in many settings, this constraint can limit flexibility in controlling attention magnitudes and may contribute to overly concentrated or unstable attention patterns during training. Prior work has explored modifications such as attention sinks or gating mechanisms, but these approaches provide only limited or indirect control over attention reweighting. We propose Affine-Scaled Attention, a simple extension to standard attention that introduces input-dependent scaling and a corresponding bias term applied to softmax-normalized attention weights. This design relaxes the strict normalization constraint while maintaining aggregation of value representations, allowing the model to adjust both the relative distribution and the scale of attention in a controlled manner. We empirically evaluate Affine-Scaled Attention in large-scale language model pretraining across multiple model sizes. Experimental results show consistent improvements in training stability, optimization behavior, and downstream task performance compared to standard softmax attention and attention sink baselines. These findings suggest that modest reweighting of attention outputs provides a practical and effective way to improve attention behavior in Transformer models.