ICML 2026 Papers

Inferring molecular structures from multimodal spectroscopic measurements requires integrating complementary yet highly heterogeneous signals. However, the common paradigm of directly concatenating multispectral sequences can exhibit anomalous performance degradation, primarily due to pronounced heterogeneity and the resulting multimodal imbalance across modalities. As a remedy, we propose MM-Spectrum, a sparse Mixture-of-Experts framework tailored for multimodal multispectral spectra-to-structure elucidation. To better match the information characteristics under multispectral imbalance, **MM-Spectrum** introduces an explicit modality-aware routing mechanism that exposes spectral identity to the router in addition to token content representations. Moreover, it incorporates shared and interaction experts, together with heterogeneous expert capacities, to extract multispectral modality-unique and cross-modal synergistic information while suppressing noise-induced interference. Across full-modality, bimodal, and missing-modality settings on molecular structural elucidation, MM-Spectrum achieves consistent and substantial improvements, supported by ablation studies and interpretability analyses. Code is available at https://github.com/HHHTTY/MM-Spectrum.

Repairing system crashes discovered by kernel fuzzers like Syzkaller is a critical yet underexplored challenge in software engineering. While recent works have introduced Large Language Model (LLM) based agents for Linux kernel crash-resolution, their evaluation benchmarks are usually static and thus, do not capture the evolving nature of the Linux kernel, and suffer from potential data contamination due to LLM knowledge cutoffs. To address the above problem, we present (i) Live-kBench, an evaluation framework for self-evolving benchmarks that continuously scrapes and evaluates agents on freshly discovered kernel bugs, and (ii) kEnv, an agent-agnostic standardized crash-resolution environment for kernel compilation, execution, and feedback. This design decouples agent workflows from heavy-weight execution, enabling fair and scalable comparison across diverse agent frameworks under identical conditions. To this end, we curate an inaugural dataset of 534 Linux kernel bugs and empirically demonstrate a significant performance gap, with agents achieving up to 25% higher equivalent patch rate on bugs fixed before the LLM knowledge cutoff. Using kEnv, we benchmark three state-of-the-art agents, showing that they resolve 74% of crashes on the first attempt (plausible patches); however only ~20% of generated patches closely match developer fixes. Additionally, exposing crash resolution feedback improves crash resolution rate by 29%. Live-kBench provides the community with an evaluation infrastructure for self-evolving benchmarks that is both time and attribute sensitive; complete with a public dashboard to track agent progress on Linux kernel bugs.

We study one-sided and $\alpha$-correct sequential hypothesis testing for data generated by an ergodic, finite-state Markov chain. The *null* hypothesis is that the unknown transition matrix belongs to a prescribed set $\cal P$ of stochastic matrices, and the *alternative* corresponds to a disjoint set $\cal Q$. We establish *a non-asymptotic instance-dependent lower bound* on the expected stopping time of any valid sequential test under the alternative, which is asymptotically tight. Our novel analysis improves the existing lower bounds, which are either asymptotic or provably sub-optimal in this setting. Our lower bound incorporates both the stationary distribution and the transition structure induced by the unknown Markov chain. We further propose an optimal test whose expected stopping time matches this lower bound asymptotically as $\alpha \to 0$. We illustrate the usefulness of our framework through applications to sequential detection of model misspecification in Markov Chain Monte Carlo and to testing structural properties, such as the linearity of transition dynamics, in Markov decision processes. Our findings yield a sharp and general characterization of optimal sequential testing procedures under Markovian dependence.

Aligning large language models (LLMs) with human preferences is commonly done via reinforcement learning from human feedback (RLHF) with Proximal Policy Optimization (PPO) or, more simply, via Direct Preference Optimization (DPO). While DPO is stable and RL-free, it treats preferences as flat winner vs. loser signals and is sensitive to noisy or brittle preferences arising from fragile chains of thought. We propose TUR-DPO, a topology- and uncertainty-aware variant of DPO that rewards how answers are derived, not only what they say, by eliciting lightweight reasoning topologies and combining semantic faithfulness, utility, and topology quality into a calibrated uncertainty signal. A small learnable reward is factorized over these signals and incorporated into an uncertainty-weighted DPO objective that remains RL-free and relies only on a fixed or moving reference policy. Empirically, across open 7–8B models and benchmarks spanning mathematical reasoning, factual question answering, summarization, and helpful/harmless dialogue, TUR-DPO improves judge win-rates, faithfulness, and calibration relative to DPO while preserving training simplicity and avoiding online rollouts. We further observe consistent gains in multimodal and long-context settings, and show that TUR-DPO matches or exceeds PPO on reasoning-centric tasks while maintaining operational simplicity.

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Zeroth-order optimization has emerged as a promising approach for fine-tuning large language models on domain-specific data, particularly under differential privacy (DP) and memory constraints. While first-order methods have been extensively studied from a privacy perspective, the privacy analysis and algorithmic design for zeroth-order methods remain significantly underexplored. A critical open question concerns hidden-state DP analysis: although convergent privacy bounds are known for first-order methods, it has remained unclear whether similar guarantees can be established for zeroth-order methods. In this work, we provide an affirmative answer by proving a convergent DP bound for zeroth-order optimization. Our analysis generalizes the celebrated privacy amplification-by-iteration framework to the setting of smooth loss functions in zeroth-order optimization. Furthermore, it induces better DP zeroth-order algorithmic designs that are previously unknown to the literature.

When computing gradients, deep learning training algorithms typically treat the mini-batch as a fundamental unit --- only returning batch-averaged gradients. Computing non-linear statistics of the mini-batch gradient distribution has traditionally been viewed as prohibitively expensive or requiring complex, custom implementations. We challenge this view by demonstrating that sequence-level architectures offer a natural testbed for prototyping algorithms based on per-example gradients. We show that staged programming languages like JAX enable generic manipulations of mini-batch gradient computations. We then build on Dangel et. al. (2019) to derive implementations of specific per-example or per-token operations with negligible computational or memory overhead. Finally, we leverage our findings to re-examine two nonlinear optimization operations. First, we analyze signSGD, showing that the optimal placement of the sign operation is critical to success and can be predicted via a simple signal-to-noise ratio argument. Second, we investigate per-example variations of the Adam preconditioner and find that, contrary to conventional wisdom, optimization is best served when the preconditioner is dominated by the mean squared of the gradient distribution rather than its variance. Overall our work shows that accessible per-example gradient information unlocks new avenues for algorithm analysis and design.

Diffusion models have shown remarkable empirical success in sampling from rich multi-modal distributions. Their inference relies on numerically solving a certain differential equation. This differential equation cannot be solved in closed form, and its resolution via discretization typically requires many small iterations to produce *high-quality* samples. More precisely, prior works have shown that the iteration complexity of discretization methods for diffusion models scales polynomially in the ambient dimension and the inverse accuracy $1/\varepsilon$. In this work, we propose a new solver for diffusion models relying on a subtle interplay between low-degree approximation and the collocation method, and we prove that its iteration complexity scales *polylogarithmically* in $1/\varepsilon$, yielding the first "high-accuracy" guarantee for a diffusion-based sampler that only uses (approximate) access to the scores of the data distribution. In addition, our bound does not depend explicitly on the ambient dimension; more precisely, the dimension affects the complexity of our solver only through the *effective radius* of the support of the target distribution.

Generalized Category Discovery (GCD) assigns unlabeled instances, mixed with labeled data, to known or novel categories, requiring human-like compositional reasoning: reusing primitives learned from known classes and deciding when new combinations imply new categories. Existing GCD methods operate on unstructured token features and struggle to extrapolate to novel compositions. We propose CoGe-GCD, which rethinks GCD through compositional generalization with two coupled stages. (i) Compositional Perception structures patch tokens by mapping them to a small vocabulary of primitives and refining token embeddings via competitive token-primitive assignment and information passing, yielding coherent groups for discovery. (ii) Generalizing Induction exploits the induced geometric structure and applies a structure-preserving calibration over spatial relations, maintaining probabilistic semantics while improving extrapolation to unseen primitive combinations. CoGe-GCD is implemented as an inductive-bias module between backbone and projection head, without modifying heads or losses, and can be plugged into diverse GCD frameworks. On standard benchmarks, it consistently improves all-class accuracy, unknown-class number estimation, and geometric quality, with marginal computational overhead.

Evaluating survival models under censoring is inherently challenging, yet standard evaluation practices are often applied without explicitly assessing how censoring distorts metric reliability. Performing a large experimental study, we analyze and quantify how survival evaluation metrics are affected in fundamentally different ways by the censoring rate and the censoring mechanism. Using a controlled semi-synthetic framework, we vary both the censoring mechanism (administrative, independent, covariate-dependent) and the censoring rate, and compare standard evaluations based on censored data with oracle evaluations using fully observed event times. This controlled setting enables us to quantify distortions along two complementary axes: numerical bias and preservation of model ranking. Across datasets and metric families, we find that censoring induces systematic, mechanism-dependent distortions. Moderate numerical bias, if not properly addressed, can lead to unreliable model comparison as censoring increases. These findings reveal fundamental limitations of common benchmarking practices and call for more careful interpretation of survival evaluation under realistic censoring.

As Large Language Models (LLMs) are increasingly deployed in social and strategic scenarios, it becomes critical to understand where and why their behavior diverges from that of humans. While behavioral game theory (BGT) provides a framework for analyzing behavior, existing models do not fully capture the idiosyncratic behavior of humans or black-box, non-human agents like LLMs. We employ AlphaEvolve, a cutting-edge program discovery tool, to directly discover *interpretable* models of human and LLM behavior from data, thereby enabling open-ended discovery of structural factors driving human and LLM behavior. Our analysis on iterated rock-paper-scissors reveals that frontier LLMs can be capable of deeper strategic behavior than humans. These results provide a foundation for understanding structural differences driving differences in human and LLM behavior in strategic interactions.

Learning-based Traffic Signal Control (TSC) achieves satisfactory performance in small networks, but its effectiveness often deteriorates in larger networks under dynamic traffic patterns and intersection heterogeneity. In this work, we propose SLight, a policy-aware grouped MARL-TSC framework that enables scalability and efficiency balance under dynamic and heterogeneous traffic conditions. SLight captures policy-influenced traffic patterns with a policy-aware traffic pattern encoder, learns explicit group-level shared control principles from state–action trajectories, and matches each intersection’s traffic pattern embedding to principle prototypes flexibly through a compatibility-based adaptive assignment module. Experiments on real-world and synthetic networks demonstrate that SLight sustains performance gains as scale increases and outperforms existing rule-based, reinforcement learning, and grouping-based baselines. Code is available at \url{https://github.com/MaHaoLun/Slight-code.git}

While neuroscience experiments have repeatedly demonstrated the involvement of large populations of neurons in even simple behaviors, these studies have just as often reported that the collective dynamics of neural activity are approximately low-dimensional. As a result, methods for identifying low-dimensional latent representations of time series data have become increasingly prominent in neuroscience. However, most existing methods either ignore temporal structure or model time evolution using latent dynamical systems approaches. In the first case, dynamics may be distorted or even scrambled in the latent space, while in the second, many possible latent dynamics may give rise to the same data. Here, we address these challenges using a novel flow-matching approach in which data are generated by a pair of flow fields, one governing time evolution, the other a mapping between data and a low-dimensional latent space. Importantly, the dimension-reducing flow is trained to minimize distortions of the temporal dynamics, learning an identifiable low-dimensional representation that preserves temporal relations in the original data. Additionally, we constrain our latent spaces to have low-dimensional support in a soft, parameterized manner, taking inspiration from ideas on nested dropout. Across both neural and behavioral data, we show that this dual flow approach produces both more interpretable dynamics and higher-quality reconstructions than competing models, including in noise-dominated data sets where conventional approaches fail.

Deep stochastic processes have recently become a central paradigm for image enhancement, with many methods explicitly conditioning the stochastic trajectory on the degraded input. However, the relationship between these conditional processes and standard diffusion models remains unclear. In this work, we introduce a unified perspective on stochastic image enhancement by classifying recent methods into three families of continuous-time processes: unconditional diffusion models, Ornstein–Uhlenbeck (OU) processes, and diffusion bridges. We show that all of these approaches arise from a common stochastic differential equation (SDE) formulation. This framework makes explicit that seemingly disparate methods differ primarily in their drift and diffusion terms, terminal distributions, and boundary conditions, while schedulers and samplers constitute orthogonal design choices. Leveraging this unification, we conduct a controlled empirical study across multiple image enhancement tasks using identical architectures and training protocols. Our results reveal no consistently dominant method; instead, we identify and disentangle the specific design choices that most strongly influence performance. Finally, we release ItoVision, a modular PyTorch library that implements the unified framework and enables rapid prototyping and fair comparison of stochastic image enhancement methods.

With the widespread application of data-driven classifiers in high-risk domains, group fairness has become an important research focus. However, most existing methods rely on model constraints or data reweighting, which may limit interpretability or distort the original data distribution. Granular-ball computing (GBC), as a structured and interpretable learning framework, provides a natural foundation for incorporating group fairness into data partitioning. Based on this insight, we propose a $\textbf{Fair}$ $\textbf{G}$ranular-$\textbf{B}$all $\textbf{G}$eneration framework (FairGBG), which employs fair clustering to maintain balanced proportions of sensitive groups within each granular ball (GB), thereby enhancing within-ball group fairness. Theoretical analysis shows that FairGBG can preserve high GB purity while satisfying group fairness requirements. Furthermore, we introduce a $\textbf{Fair}$ $\textbf{G}$ranular-$\textbf{B}$all-based data $\textbf{F}$air $\textbf{C}$lassification method (FairGBFC), which leverages fair GBs to improve classification fairness. Experiments on multiple benchmark datasets demonstrate that, compared with existing GB generation methods, FairGBG can generate locally fair GBs. Moreover, compared with state-of-the-art fairness-aware baselines, FairGBFC achieves a superior trade-off between accuracy and fairness, effectively mitigating bias while preserving high utility. Our code is publicly available at https://github.com/jamesandai/FairGB.

Large Vision-Language Models (LVLMs) demonstrate impressive multimodal capabilities, yet suffer from hallucination—generating factually inaccurate content. Contrastive Decoding (CD) mitigates this by contrasting amateur and expert branches at the logit level. However, our investigation reveals that such logit-level interventions fundamentally compromise generation coherence, necessitating restrictive penalty constraints unrelated to hallucination suppression. We introduce Attention Contrastive Decoding (ACD), a training-free plug-in that complements logit-level CD by relocating part of the contrastive operations to the attention mechanism. Operating at an earlier stage of the forward pass, ACD performs smooth semantic-preserving interventions through an Adaptive Subtraction Strategy (ASS), which attenuates hallucination-associated attention patterns while amplifying critical visual information. Extensive experiments demonstrate that combining ACD with existing CD methods (e.g., VCD+ACD) produces substantially more coherent outputs with further reduced hallucinations, eliminating restrictive penalties while enabling trustworthy multimodal generation.

Large language models (LLMs) have recently shown strong performance on mathematical benchmarks. At the same time, they are prone to hallucination and sycophancy, often providing convincing but flawed proofs for incorrect mathematical statements provided by users. This significantly limits the applicability of LLMs in theorem proving, as verification of these flawed proofs must be done manually by expert mathematicians. However, existing benchmarks that measure sycophancy in mathematics are limited: they focus solely on final-answer problems, rely on very simple and often contaminated datasets, and construct benchmark samples using synthetic modifications that create ill-posed questions. To address these issues, we introduce BrokenMath, the first benchmark for evaluating sycophantic behavior in LLMs within the context of natural language theorem proving. BrokenMath is built from advanced 2025 competition problems, which are perturbed with an LLM to produce false statements and subsequently refined through expert review. We evaluate state-of-the-art LLMs and agentic systems and find that sycophancy is widespread, with the best model, GPT-5, producing sycophantic answers 29% of the time. We further investigate several mitigation strategies, including test-time interventions and supervised fine-tuning on curated sycophantic examples. These approaches reduce, but do not eliminate, sycophancy.

Retrieval-Augmented Generation (RAG) augments large language models with external knowledge, which in turn exposes their retrieval corpora to data poisoning risks. However, existing poisoning attacks exhibit limited effectiveness against RAG equipped with a reranker to enhance retrieval quality. Remarkably, this defensive capability requires no adversarial training: a reranker fine-tuned solely on benign, in-domain corpora can effectively filter malicious content. Towards realistic RAG red-teaming, we conclude practical prompt design principles that reveal reranker blind spots. Building on these insights, we introduce the Prompt-Perturbation Poisoning Attack ($\mathbf{P}^3 \mathbf{A}$). $\mathbf{P}^3 \mathbf{A}$ first employs rule-based prompt engineering to craft initial poisoned texts. It then injects subtle character-level perturbations into these texts, which promotes their ranking by the reranker while maintaining their adversarial effectiveness. These perturbations introduce only about 1\% textual change, ensuring the poisoned texts remain natural and readable. Extensive experiments show that $\mathbf{P}^3 \mathbf{A}$ achieves strong attack effectiveness and transferability, even when constrained to poisoning a single document. Code is available at https://github.com/YyyxKun/P3A.

Reinforcement learning has emerged as a principled post-training paradigm for Temporal Video Grounding (TVG) due to its on-policy optimization, yet existing GRPO-based methods remain fundamentally constrained by sparse reward signals and substantial computational overhead. We propose Video-OPD, an efficient post-training framework for TVG inspired by recent advances in on-policy distillation. Video-OPD optimizes trajectories sampled directly from the current policy, thereby preserving alignment between training and inference distributions, while a frontier teacher supplies dense, token-level supervision via a reverse KL divergence objective. This formulation preserves the on-policy property critical for mitigating distributional shift, while converting sparse, episode-level feedback into fine-grained, step-wise learning signals. Building on Video-OPD, we introduce Teacher-Validated Disagreement Focusing (TVDF), a lightweight training curriculum that iteratively prioritizes trajectories that are both teacher-reliable and maximally informative for the student, thereby improving training efficiency. Empirical results demonstrate that Video-OPD consistently outperforms GRPO while achieving substantially faster convergence and lower computational cost, establishing on-policy distillation as an effective alternative to conventional reinforcement learning for TVG.