ICML 2026 论文列表

The advancement of Large Language Models (LLMs) into higher-level reasoning domains has rendered traditional heuristic evaluators insufficient for long-form open-ended responses, precipitating the widespread adoption of rubric-based benchmarks. While these frameworks utilize expert-curated criteria and LLM-as-a-judge to assess open-ended generation, the intrinsic robustness of these evaluation harnesses to fundamental validity assessments remains critically under-investigated. To bridge this gap, we introduce RubricRobustness, a systematic sensitivity analysis framework that subjects these benchmarks to three common sense perturbations: *semantic negation*, *stochastic deletion* and *irrelevant addition*. We investigate the extent to which manipulating the semantic veracity of a model's response impacts its resulting score by applying the robustness framework to two of the most popular rubrics-based benchmarks: HealthBench and WildBench. Our findings reveal systematic vulnerabilities: while both benchmarks respond sharply to semantic negation (e.g., degradation slopes of approximately $-0.38$ on HealthBench and $-0.55$ on WildBench), they are substantially less responsive to irrelevant addition, often requiring over 35% of sentences to be perturbed before inducing even a 25% score drop. We argue that perturbation-based sensitivity analyses of this form are a necessary prerequisite for validating rubric coverage, ensuring that automated evaluation frameworks reliably penalize basic semantic failures. We will release our framework as an open-source tool for building more resilient benchmarks.

Zeroth-order optimization is widely used when gradients are unavailable, but the standard two-point estimator suffers from $\mathcal{O}(r^2)$ truncation bias at smoothing radius $r$. Existing bias-reduction schemes typically increase the leading directional variance under a fixed number of function evaluations per gradient estimate, while variance-reduction schemes generally do not improve the bias order. We propose Romberg-ZOGE, which forms a Romberg-extrapolated linear combination of two-point differences evaluated at radii $\{r/2^k\}_{k=0}^R$ while reusing the same perturbation direction across all radii. With appropriately chosen weights, Romberg-ZOGE cancels the first $R$ even-order truncation terms and achieves $\mathcal{O}(r^{2R+2})$ bias under $(2R{+}2)$-order smoothness, while preserving the leading directional variance constant of the two-point estimator up to higher-order residual terms. We further characterize the stochastic-oracle setting by deriving an explicit noise-amplification factor and corresponding bias and variance bounds. Experiments on synthetic benchmarks, simulator-based wireless optimization, and black-box prompt tuning of OPT-1.3B demonstrate faster and more stable zeroth-order SGD when the number of function evaluations per gradient estimate is fixed.

While Physics-Informed Neural Networks (PINNs) are powerful for solving Partial Differential Equations (PDEs), their training is often paralyzed by gradient pathology. The gradients from the PDE residuals and boundary constraints oppose each other, trapping the model in local minima. Current solutions, such as adaptive weighting or hard constraints, either fail to fundamentally resolve this ill-conditioning or are limited to simple geometries. In this study, we systematically analyze the possible causes of this gradient pathology from the perspectives of loss landscapes and optimization dynamics. Based on the obtained conclusion, we propose Constraint-Aligned loss with Manifold Lifting (CAML). By reformulating all zeroth-order terms into aligned constraints, our method effectively mitigates gradient conflicts. In addition, we introduce a delay factor to help the optimizer skip the high-curvature area. Experiments demonstrate that our CAML significantly enhances numerical stability and efficiency in highly complex PINN problems. Our code is open-sourced on [CAML](https://github.com/YichenLuo-0/CAML).

We address the problem of causal effect estimation in the presence of hidden confounders using nonparametric instrumental variable (IV) regression. An established approach is to use estimators based on learned \emph{spectral features}, that is, features spanning the top singular subspaces of the operator linking treatments to instruments. While powerful, such features are agnostic to the outcome variable. Consequently, the method can fail when the true causal function is poorly represented by these dominant singular functions. To mitigate, we introduce Augmented Spectral Feature Learning, a framework that makes the feature learning process outcome-aware. Our method learns features by minimizing a novel contrastive loss derived from an augmented operator that incorporates information from the outcome. By learning these task-specific features, our approach remains effective even under spectral misalignment. We provide a theoretical analysis of this framework and validate our approach on challenging benchmarks.

Diffusion-based image synthesis has made AI-generated images (AIGI) increasingly photorealistic, raising urgent concerns about authenticity in applications such as misinformation detection, digital forensics, and content moderation. Despite the substantial advances in AIGI detection, how to correct detected AI-generated images with visible artifacts and restore realistic appearance remains largely underexplored. Moreover, few existing work has established the connection between AIGI detection and artifact correction. To fill this gap, we propose GenShield, a unified autoregressive framework that jointly performs explainable AIGI detection and controllable artifact correction in a closed loop from diagnosis to restoration, revealing a mutually reinforcing relationship between these two tasks. We further introduce a Visual Chain-of-Thought based curriculum learning strategy that enables self-explained, multi-step "diagnose-then-repair" correction with an explicit stopping criterion. A high-quality dataset with large-scale "artifact-restored" pairs is also constructed alongside a unified evaluation pipeline. Extensive experiments on our correction benchmark and mainstream AIGI detection benchmarks demonstrate state-of-the-art performance and strong generalization of our method.

Despite advances in Preference Alignment (PA) for Large Language Models (LLMs), mainstream methods like reinforcement learning with human feedback face notable challenges. These approaches require high-quality datasets of positive preference examples, which are costly to obtain and computationally intensive. The LLM unlearning technique presents a promising alternative by directly removing the influence of negative examples. However, current research has primarily focused on empirical validation, lacking systematic quantitative analysis. To bridge this gap, we propose a framework linking PA with LLM unlearning. Through bi-level optimization, we first quantify how unlearning specific negative examples impacts PA performance. Our analysis reveals that these effects vary substantially across negative examples. Building on this insight, we pose a crucial question: how can we optimally select and weight negative examples for unlearning to maximize PA performance? To answer this, we propose Unlearning to Align (U2A), which leverages bi-level optimization to efficiently select and unlearn examples for optimal PA performance. We validate the proposed method through extensive experiments, with results confirming its effectiveness. Our code is available at https://anonymous.4open.science/r/U2A-9E75.

We study optimal experimental design for multinomial logit (MNL) bandits, where an agent repeatedly selects a subset of $K$ items from a ground set of size $N$ and observes single-choice feedback. Unlike linear or generalized linear bandits, MNL bandits have a combinatorial action space, which makes classical optimal design approaches and naive optimization over all subsets computationally intractable. We propose a computationally efficient optimal design framework for MNL models that achieves both statistical efficiency and scalability through two complementary approaches: (i) an exact or certified-approximate reformulation of the design oracle as a $0$-$1$ mixed-integer linear program (MILP) with solver-certified early stopping, and (ii) a fully polynomial-time lifted design that replaces the nonlinear objective with a tractable surrogate. Using the Kiefer-Wolfowitz equivalence theorem, we establish near G-optimality guarantees and characterize the induced statistical-computational trade-offs. As an application, we develop a best assortment identification algorithm for MNL bandits with linear utilities and non-uniform revenues, and prove an instance-dependent sample complexity of $\tilde{\mathcal{O}}\big(\frac{d \log N}{\Delta^2}\big)$, where $d$ is the feature dimension, $N$ is the number of arms, and $\Delta$ is the minimum revenue gap.

Diffusion Large Language Models (dLLMs) enable parallel token generation, and their block-wise variants have attracted significant attention. However, existing dLLMs usually exhibit an accuracy–parallelism trade-off, where raising tokens per forward (TPF) via aggressive parallel decoding often degrades task accuracy. To address this, we suggest developing a post-training approach to directly optimize the speed–quality frontier of pre-trained dLLMs. Conceptually, we do not require the model to decode aggressively along all sampling trajectories, but rather to find several highly parallelizable ones that can yield correct results. To this end, we resort to a reinforcement learning paradigm, i.e., LightningRL, to optimize rewards regarding both the final accuracy and inference parallelism. LightningRL follows the Group Relative Policy Optimization (GRPO) framework, with further improvements for dLLMs: 1) stabilized training via per-reward decoupled normalization, 2) token-level negative log-likelihood (NLL) loss on correct trajectories for regularization, and 3) improved training efficiency through dynamic sampling with TPF-aware filtering. Across maths and code tasks, LightningRL consistently advances the Pareto frontier, maintaining competitive accuracy while increasing parallelism to an average TPF of 7.3 (up to 11.10 on MBPP).

The widespread adoption of text-to-image (T2I) diffusion models has raised concerns about their potential to generate copyrighted, inappropriate, or sensitive imagery. As a practical solution, *machine unlearning* aims to erase unwanted concepts without retraining from scratch. While most existing methods are effective for single-concept unlearning, they often struggle when removing multiple concepts, causing significant challenges in unlearning effectiveness, generation quality, and sensitivity to hyperparameters and datasets. We take a unique perspective on multi-concept unlearning by leveraging model sparsity and propose the **F**orget **I**t **A**ll (FIA) framework. FIA first introduces *Contrastive Concept Saliency* to quantify each weight connection's contribution to a target concept. It then identifies *Concept Sensitive Neurons* by combining temporal and spatial information, ensuring that only neurons consistently responsive to the target concept are selected. Finally, FIA constructs masks from the identified neurons and fuses them into a unified multi-concept mask, where *Concept Agnostic Neurons* that broadly support general content generation are preserved while concept-specific neurons are pruned to remove the targets. FIA is training-free and requires minimal hyperparameter tuning for new tasks, enabling plug-and-play use. Extensive experiments across three distinct unlearning tasks demonstrate that FIA achieves more reliable multi-concept unlearning, improving forgetting effectiveness while maintaining generation fidelity and quality. Code is available at [https://github.com/kaiyuan02415/Forget-It-All](https://github.com/kaiyuan02415/Forget-It-All)

Vision-Language-Action (VLA) models are promising for generalist robot control, but on-robot deployment is bottlenecked by real-time inference under tight cost and energy budgets. Most prior evaluations rely on desktop-grade GPUs, obscuring the trade-offs and opportunities offered by heterogeneous edge accelerators (GPUs/XPUs/NPUs). We present a systematic framework for low-cost VLA deployment via model--hardware co-characterization. First, we build a cross-accelerator leaderboard and evaluate model--hardware pairs under \textbf{CET} (Cost, Energy, Time), showing that ``right-sized'' edge devices can be more cost-/energy-efficient than flagship GPUs while meeting control-rate constraints. Second, using fine-grained SM tracing and Roofline analysis, we uncover a consistent two-phase inference pattern: a compute-bound VLM backbone followed by a memory-bound Action Expert, which induces phase-dependent underutilization and hardware inefficiency. Finally, guided by these insights, we propose \textbf{DP-Cache} and \textbf{V-AEFusion} to reduce diffusion redundancy and enable asynchronous pipeline parallelism, achieving up to \(2.9\times\) speedup on GPUs and \(6\times\) on edge NPUs with only marginal success degradation. The example leaderboard website is: \url{https://vla-leaderboard-01.vercel.app/}.

In reinforcement learning, human preference feedback is emerging as a viable alternative to expert-designed reward functions, which can be difficult to engineer in real-world problems. However, despite the growing importance of preference feedback, how to effectively *elicit* preferences remains a fundamental open problem. This work focuses on the *compatible reward identification* task. The aim is to derive, starting from preference feedback, a reward function compatible with the observed preferences and accurate across the *entire* state-action space, ensuring higher transferability, safety, and interpretability. Indeed, the most common *reinforcement learning from human feedback* objective is to learn the *optimal policy*, requiring accuracy only in the portion of the state-action space that the agent visits. However, this goal cannot provide the same guarantees as compatible reward identification. First, we discuss commonalities and differences between the two goals. Then, we consider deterministic preferences, deriving the minimum number of interactions needed to identify the set of compatible rewards, and showing that using fewer queries may lead to arbitrarily large suboptimality. Finally, we focus on stochastic preferences generated via the Bradley-Terry (BT) model. We introduce the concepts of *query basis* and its *index*, relating them to the problem complexity. Upon this, we discuss the connection between the index of a basis and the BT model, as well as the limitations that the model induces in this setting. Additionally, we devise an algorithm to identify a nearly-optimal query basis with polynomial human query complexity.

Unified audio--visual generation is rapidly gaining industrial and creative relevance, enabling applications in virtual production and interactive media. However, when moving from general audio--video synthesis to music–dance co-generation, the task becomes substantially harder: musical rhythm, phrasing, and accents must drive choreographic motion at fine temporal resolution, and such rhythmic coupling is not captured by unimodal metrics or generic audiovisual consistency scores used in current evaluation practice. We introduce TMD-Bench, a benchmark for text-driven music–dance co-generation that assesses systems across unimodal generation quality, instruction adherence, and cross-modal rhythmic alignment. The benchmark integrates computable physical metrics with perceptual multimodal judgments, and is supported by a curated rhythm-aligned music–dance dataset and a fine-grained Music Captioner for structured music semantics. TMD-Bench further reveals that (i) modern commercial audio--visual models (e.g., Veo 3, Sora 2) produce high-quality music and video, while rhythmic coupling remains less consistently optimized and leaves room for improvement, and (ii) our unified baseline RhyJAM trained on rhythm-aligned data achieves competitive beat-level synchronization while maintaining competitive unimodal fidelity. This presents prospects for building next-generation music–dance models that explicitly optimize rhythmic and kinetic coherence. The code is available at https://github.com/MM-Speech/TMD-Bench.

Modern visual object trackers show impressive results on general targets, yet their performance drops substantially when dealing with scene text. Although currently underexplored, tracking text in videos is essential for dynamic text manipulations such as segmentation, removal, and editing. To fill this gap, this paper formalizes this specific task as Scene Text Tracking and presents the first systematic work for it. We identify three primary challenges in this task: 1) severe geometric distortions from perspective shifts, 2) high visual ambiguity across different instances, and 3) high sensitivity to fine-grained structural details. To address these issues, we propose SymTrack, a unified detection-free framework with synergistic dual-branch design. It integrates a Cross-Expert Calibration mechanism to reduce semantic bias, along with a Predictive Token Rectification mechanism to correct structural imbalances, complemented by an Adaptive Inference Engine that stabilizes predictions under motion constraints. Considering the lack of dedicated benchmarks for this task, we utilize three datasets from video text spotting to construct a benchmark with high-quality annotations. Extensive experiments demonstrate that SymTrack sets the new state-of-the-art on all three benchmarks, outperforming previous best trackers by up to 11.97\% AUC on $ \text{BOVText}_{\text{SOT}} $. Overall, our work promotes efficient and thorough text tracking, paving the way toward more generalized video text manipulation.

Belief Propagation (BP) is a powerful algorithm for distributed inference in probabilistic graphical models, however it quickly becomes infeasible for practical compute and memory budgets. Many efficient, non-parametric forms of BP have been developed, but the most popular is Gaussian Belief Propagation (GBP), a variant that assumes all distributions are locally Gaussian. GBP is widely used due to its efficiency and empirically strong performance in applications like computer vision or sensor networks – even when modelling non-Gaussian problems. In this paper, we seek to provide a theoretical guarantee for when Gaussian approximations are valid in highly non-Gaussian, sparsely-connected factor graphs performing BP (common in Spatial AI). We leverage the Central Limit Theorem to prove mathematically that variables’ beliefs under BP converge to a Gaussian distribution in complex, loopy factor graphs obeying our 4 key assumptions. We then confirm experimentally that variable beliefs become increasingly Gaussian after just a few BP iterations in a stereo depth estimation task.

GPT-style language models are sensitive to single-token changes at generation points where the predicted probability distribution is spread across multiple tokens. Viewing this sensitivity as a geometric property, we derive an $\mathfrak{so}(n)$-valued 1-form that depends only on the geometry of the token embeddings. Despite this purely geometric origin, we show that its curvature is semantically meaningful: on chess reasoning tasks, the curvature couples to the world model of an off-the-shelf instruction-tuned model, with transformations clustering by board region and respecting piece importance. Our findings suggest that token space geometry directly reflects how models internally represent problems.

Recent advancements in Video Large Language Models (Video LLMs) have demonstrated impressive results, yet existing approaches handle either temporal or spatial dimension in isolation, struggling in the analysis of complex events that require spatial-temporal integration. To bridge this gap, we propose VideoLoom, a unified Video LLM for joint spatial-temporal understanding. To facilitate the development of fine-grained spatial and temporal localization capabilities, we curate LoomData-8.7k, a character-centric video dataset with temporally grounded and spatially localized captions. With this, VideoLoom achieves the state-of-the-art performance across a variety of spatial and temporal benchmarks. In addition, we introduce LoomBench, a benchmark consisting of temporal, spatial, and compositional video–question pairs, with a novel metric $J$&$F_{bi-fore}$, enabling a comprehensive evaluation of Video LLMs from diverse aspects. Collectively, these contributions offer a universal and effective suite for joint spatial-temporal video understanding, setting a new standard in multimodal intelligence.

We investigate the distributionally robust assortment optimization (DRAO) problem under the contextual multinomial logit (MNL) choice model, where the decision-maker seeks to maximize revenue against worst-case distributional deviations. To address potential distribution shifts relative to the observed data environment, we study DRAO under ambiguity sets defined by three divergences: total variation (TV), Kullback–Leibler (KL), and chi-square ($\chi^2$). Incorporating robust concerns poses challenges for both algorithm design and theoretical analysis. By leveraging strong duality results from the distributionally robust optimization literature and integrating them into the assortment optimization procedures, we develop tailored polynomial-time algorithms under each divergence. We further provide a theoretical analysis and establish sample complexity bounds for all three robust formulations.

Probabilistic forecasting with exogenous variables is vital for decision-making but remains underexplored compared to deterministic methods. We propose KITE, a knowledge-guided probabilistic modeling framework designed to bridge this gap by addressing two key bottlenecks: (1) topological disparity in sampling initialization and (2) spurious covariate correlations during the iterative conditional generation process. KITE introduces a *History-Conditional Manifold* to construct an informative source distribution from historical dynamics, effectively anchoring the starting point closer to the target space. Additionally, a *Knowledge-Guided Conditioning* module is developed to regularize variable interactions using statistical priors, suppressing spurious correlations and enhancing the robustness of covariate conditioning. Extensive experiments demonstrate that KITE outperforms state-of-the-art methods in both deterministic and probabilistic forecasting.

Being able to edit panoramic images is crucial for creating realistic 360° visual experiences. However, existing perspective-based image editing methods fail to model the spatial structure of panoramas. Conventional cube-map decompositions attempt to overcome this problem but inevitably break global consistency due to their mismatch with spherical geometry. Motivated by this insight, we reformulate panoramic editing directly in the equirectangular projection (ERP) domain and present World Shaper, a unified geometry-aware framework that supports five distinct editing operations within a single ERP-native representation. To address the latitude-dependent geometric distortion inherent in ERP, we introduce a geometry-aware learning strategy comprising distortion-aware attention modulation (DAAM), which steers cross-attention with latitude-dependent strength at the feature level; layered shape loss (LSL), which enforces per-object geometric supervision at the output level; and progressive curriculum training to internalize panoramic priors. To overcome the scarcity of paired panoramic editing data, we train a dedicated ERP-native controllable generator that synthesizes objects directly in the equirectangular domain under user-defined conditions, enabling scalable paired data construction for learning diverse editing behaviors. Extensive experiments on our new benchmark, PEBench, demonstrate that World Shaper achieves superior geometric consistency, editing fidelity, and text controllability compared to state-of-the-art methods, enabling coherent and flexible 360° visual world creation with unified editing control. Code, models, and data are available at the project page: https://world-shaper-project.github.io/.