ICML 2026 论文列表

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

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

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

The additive noise mechanism is a foundational tool for differential privacy (DP) of $T$-dimensional real-valued vector queries. The Gaussian mechanism, utilizing Gaussian noise, is the mostly widely used such mechanism, due to its simplicity and strong privacy guarantees. In this work, we provide justification for this choice, showing that as the dimension $T\to\infty$, no additive-noise mechanism can asymptotically improve on the Gaussian mechanism's privacy--utility tradeoff for the strong privacy settings typically used. We also develop a new family of **Spherical Generalized Gamma** DP mechanisms, which contains both the Gaussian mechanism and the recently studied $\ell_2$ mechanism (Joseph \emph{et al.}, ICML 2025). We identify members of this family that outperform both the Gaussian and $\ell_2$ mechanisms in certain low-dimensional settings, and show tight composition of all mechanisms in this family, answering an open question of Joseph \emph{et al.}~regarding the $\ell_2$ mechanism.

Hybrid sequence models—combining Transformer and state-space model layers—seek to gain the expressive versatility of attention as well as the computational efficiency of state-space model layers. Despite burgeoning interest in hybrid models, we lack a basic understanding of the settings where—and underlying mechanisms through which—they offer benefits over their constituent models. In this paper, we study this question, focusing on a broad family of core synthetic tasks. For this family of tasks, we prove the existence of fundamental limitations for non-hybrid models. Specifically, any Transformer or state-space model that solves the underlying task requires either a large number of parameters or a large working memory. On the other hand, for two prototypical tasks within this family—namely selective copying and associative recall—we construct hybrid models of small size and working memory that provably solve these tasks, thus achieving the best of both worlds. Our experimental evaluation empirically validates our theoretical findings. Importantly, going beyond the settings in our theoretical analysis, we empirically show that learned—rather than constructed—hybrids outperform non-hybrid models with up to $6 \times$ as many parameters. We additionally demonstrate that hybrid models exhibit stronger length generalization and out-of-distribution robustness than non-hybrids.

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

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

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

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

The standard oracle model for matroid algorithms assumes that each independence query can be answered in constant time, regardless of the size of the queried set. While this abstraction has underpinned much of the theoretical progress in matroid optimization, it masks the true computational effort required by these algorithms. In particular, for natural and widely studied classes such as graphic matroids, even a single independence query can require work linear in the size of the set, making the constant-time assumption implausible. We address this gap by introducing a size-sensitive cost model where the cost of a query $Q$ scales with $|Q|$. Nearly linear-time oracle implementations exist for broad families of matroids, and this refined abstraction therefore captures the true cost of query evaluation while allowing for a more faithful comparison between general matroids and their natural special cases. Within this framework we study three fundamental algorithmic tasks: finding a basis of a matroid, approximating its rank, and approximating its partition size. We establish tight results, proving nearly matching upper and lower bounds that show the optimal query cost is (up to logarithmic factors) quadratic in the size of the matroid. On the algorithmic side, our upper bounds are realized by explicit procedures that construct the desired solution. On the complexity side, our lower bounds are unconditional and already hold even for weaker distinguishing formulations of the problems. Finally, for matroids with maximum circuit size at most $c$, we show that the quadratic barrier can be broken, providing an algorithm that calculates the maximum-weight basis with expected query cost $\mathcal{O}(n^{2-1/c} \log n)$.

The automated design of agentic systems offers a promising pathway for scaling large language models (LLMs) beyond single-agent reasoning. While prior work has advanced task performance through handcrafted or automatically generated multi-agent workflows, robustness is often treated as an afterthought, leaving systems vulnerable to external adversaries and internal failures. We propose AutoRAS, a framework for the Automated design of Robust Agentic Systems. AutoRAS formulates system design as generating a sequence of symbolic primitives that jointly encode structural connectivity and behavioral actions, and learns to optimize this sequence using execution-derived safety signals and flow-based sequence-level objectives. Extensive experiments show that AutoRAS achieves the best performance in both vanilla and adversarial settings, with the smallest performance degradation under attacks. Further analyses demonstrate strong transferability, stable optimization behavior, stability across primitive sets, and favorable cost trade-offs. Our code is available at [this link](https://github.com/guohezuy/AutoRAS).

Mixture-of-Experts (MoE) models scale by activating only a small subset of experts per token. However, training such models remains challenging because top-$k$ routing is discrete and non-differentiable, requiring gradient estimators for expert selection whose design remains a central open problem. We introduce ProbMoE, a probabilistic routing framework that models expert selection as a distribution over cardinality-constrained expert subsets and formulates routing as probabilistic inference in this discrete subset space. We first propose ProbMoE Exact-$k$ routing, which samples $k$-expert subsets in the forward pass, and the backward pass uses gradients through each expert's exact marginal probability as a tractable surrogate for the true gradient. ProbMoE naturally generalizes to a dynamic-$k$ routing setting, where both training and inference constrain the routing cardinality to the same predefined range, allowing adaptive expert allocation per token. Across benchmarks and model backbones, ProbMoE Exact-$k$ achieves strong performance compared to competitive baselines, with improved expert utilization and routing diversity; ProbMoE Dynamic-$k$ achieves comparable performance with fewer activated experts. Code is available at: https://github.com/HengHugoZhao/ProbMoE.git

Pre-deployment evaluations inspect only a limited sample of model actions. A misaligned model could evade oversight by randomizing the timing of policy-violating actions, executing them so rarely that none are observed during evaluation. But this requires taking actions at very low rates, while maintaining calibration. Are frontier models even capable of that? We prompt the GPT-5, Claude 4.5 and Qwen-3 families to take a target action at low probabilities (e.g. 0.01%), either given directly or requiring derivation, and evaluate their calibration (i.e. whether they perform the target action roughly 1 in 10,000 times when resampling). We find that frontier models are surprisingly good at this task. If there is a source of entropy in-context (such as a UUID), they maintain high calibration at rates as low as 1 in 100,000 actions. Without external entropy, some models can still reach rates lower than 1 in 10,000. When target rates are given, larger models achieve good calibration at lower rates. Yet, when models must derive the optimal target rate themselves, all models fail to achieve calibration without entropy or hint to generate it. Successful low-rate strategies exhibit explicit Chain-of-Thought (CoT) reasoning, so misaligned models attempting this approach could currently be caught by a CoT monitor. However, scaling trends suggest future evaluations may be unable to rely on models' lack of target rate calibration, especially if CoT is no longer legible.

Large language model (LLM) agents increasingly perform commercial tasks that involve retrieving external content such as images and, where appropriate, reproducing that content. LLM agents should comply with the law, including copyright law. Presently, however, we lack adequate frameworks to assess whether they do so in practice. To that end, we introduce **Copyright-Bench**, a benchmark designed to evaluate *LLM agents' compliance with* *copyright law*. Copyright-Bench is comprised of realistic commercial tasks---website development, merchandise design, and pitch deck production---that involve agents selecting between public-domain content (the use of which is *legal*) and copyrighted content (the use of which is *infringing* in this setting). The evaluation introduces prompt variations that simulate different user preferences, as well as time pressure. Comparing state-of-the-art LLM agents against a human baseline, we find that: (1) agents select copyrighted works despite the availability of public-domain alternatives; and (2) for open-weights models, violation rates increase in response to certain user preferences and simulated time pressure.

Sharpness-Aware Minimization (SAM) is widely used to seek flatter minima—often linked to better generalization. In its standard implementation, SAM updates the current iterate using the loss gradient evaluated at a point perturbed by distance $\rho$ along the normalized gradient direction. We show that, for some choices of $\rho$, SAM can stall at points where this shifted (perturbed-point) gradient vanishes despite a nonzero original gradient, and therefore, they are not stationary points of the original loss. We call these points hallucinated minimizers, prove their existence under simple nonconvex landscape conditions (e.g., the presence of a local minimizer and a local maximizer), and establish sufficient conditions for local convergence of the SAM iterates to them. We corroborate this failure mode in neural network training and observe that it aligns with SAM's performance degradation often seen at large $\rho$. Finally, as a practical safeguard, we find that a short initial SGD warm-start before enabling SAM mitigates this failure mode and reduces sensitivity to the choice of $\rho$.

Social media platforms enable large-scale cross-lingual communication, but translating user-generated content (UGC) remains challenging due to its informal style, cultural references, and interaction-based expressions. While recent LLMs have improved translation quality, existing benchmarks and metrics often fail to capture whether translations convey intended meaning and cultural resonance in real-world settings. In this work, we introduce **CULTURE-MT**, a benchmark for social media translation that focuses on both **CUL**tural **T**ransmission and **U**GC-specific emotion **RE**sonance. CULTURE-MT consists of 1,002 UGC notes across 14 domains, categorized into four types based on culture-loaded symbol and linguistic style features. We also construct UGC-oriented training data to fine-tune Qwen3-8B and Qwen3-32B as baselines. We propose **cultural effectiveness** as a new evaluation criterion, focusing on expression accuracy and cultural adaptability. Testing 15 models, including the baselines, we find that traditional metrics fail to capture cultural effectiveness. We also observe that cultural effectiveness on base LLMs correlates with model size. Our work provides a comprehensive evaluation system for UGC translation models and will offers an open evaluation platform to advance research in this area. We release the CULTURE-MT benchmark and provide an online leaderboard where submitted translation results can be evaluated by our trained JUDGER.

Speculative decoding accelerates inference for (M)LLMs, yet a training-decoding discrepancy persists: while existing methods optimize single greedy trajectories, decoding involves verifying and ranking multiple sampled draft paths. We propose *Variational Speculative Decoding* (VSD), formulating draft training as variational inference over latent proposals (draft paths). VSD maximizes the marginal probability of target-model acceptance, yielding an ELBO that promotes high-quality latent proposals while minimizing divergence from the target distribution. To enhance quality and reduce variance, we incorporate a path-level utility and optimize via an Expectation-Maximization procedure. The E-step draws MCMC samples from an oracle-filtered posterior, while the M-step maximizes weighted likelihood using Adaptive Rejection Weighting (ARW) and Confidence-Aware Regularization (CAR). Theoretical analysis confirms that VSD increases expected acceptance length and speedup. Extensive experiments across LLMs and MLLMs show that VSD achieves up to a 9.58\% speedup over EAGLE-3 and 8.80\% over ViSpec, significantly improving decoding efficiency.

The trustworthiness of evaluation is critical to reliable model comparison and deployment in Video Anomaly Understanding (VAU). However, existing metrics are sensitive to expression styles and normal content, and this field lacks a diagnostic benchmark to validate metric validity and robustness. To bridge this gap, we propose: (1) a Class-Guided Chain-of-Evaluation (CG-CoE) metric, which structures assessment by extracting anomalous events and matching them under a class-specific semantic tolerance boundary, thereby decoupling anomaly semantics from descriptive style; and (2) an anomaly-focused meta-evaluation benchmark with two subsets: Anomalous Event-level Annotations (AEA) for measuring the validity of reflecting VAU models’ anomaly understanding ability and Controlled Variant Pairs (CVP) with fixed anomalies for quantifying robustness to stylistic perturbations. Extensive experiments demonstrate that CG-CoE achieves SOTA validity and robustness.

Optimizing data mixtures is is essential for unlocking the full potential of of large language models (LLMs), yet identifying the optimal composition remains computationally prohibitive due to reliance on heuristic trials or expensive proxy training. To address this, we introduce MergeMix, a novel appraoch that efficiently determines optimal data mixing ratios by repurposing model merging weights as a high-fidelity, low-cost performance proxy. By training domain-specific experts on minimal tokens and optimizing their merging weights against downstream benchmarks, MergeMix effectively optimizes the performance of data mixtures without incurring the cost of full-scale training. Extensive experiments on models with 8B and 16B parameters validate that MergeMix achieves performance comparable to or surpassing exhaustive manual tuning while drastically reducing search costs. Furthermore, MergeMix exhibits high rank consistency (Spearman $\rho > 0.9$) and strong cross-scale transferability, offering a scalable, automated solution for data mixture optimization.