ICML 2026 论文列表

Reliable reward models (RMs) are critical for ensuring the safe alignment of large language models (LLMs). However, current RM evaluation methods focus solely on preference perception accuracies in given specific scenarios, obscuring the critical vulnerabilities of RMs in real-world scenarios. We identify the true challenge lies in assessing a novel dimension: Suitability, defined as conditional reliability under specific real-world perturbations. To this end, we introduce Reward Auditor, a hypothesis-testing framework specifically designed for RM suitability inference. Rather than answering “How accurate is the RM's preference perception for given samples?”, it employs scientific auditing to answer: “Can we infer RMs exhibit systematic vulnerabilities in specific real-world scenarios?". Under real-world perturbed scenarios, Reward Auditor quantifies statistical significance and effect size by auditing distribution degradation of RM preference perception confidence. This enables inference of both the certainty and severity of RM vulnerabilities across diverse real-world scenarios, thereby laying a solid foundation for building next-generation LLM alignment systems that are verifiably safe, more robust, and trustworthy.

AI assistants in human-AI collaboration often correct suboptimal human actions through behavioral feedback (e.g., alerts or steering-wheel nudges in assistive driving). Such interventions can mitigate immediate errors, but long-term improvement requires addressing the underlying misconceptions that cause repeated mistakes. We introduce SENSEI, a framework that infers user misconceptions from interaction behavior and provides targeted, minimal yet sufficient suggestions to correct them. Our approach departs from action- or trajectory-level interventions by operating over a structured knowledge representation to localize and correct the sources of erroneous behavior. Across three long-horizon tasks with diverse misconceptions and corresponding behaviors, SENSEI demonstrates zero-shot compositional generalization, disentangling multiple overlapping misconceptions despite training only on single-misconception cases. A user study further shows that our method identifies real human misconceptions and provides effective guidance that improves long-horizon task performance, successfully correcting 90% of student misconceptions.

Medical vision-language models (VLMs) have rapidly advanced as general-purpose multimodal assistants, yet their deployment in 3D Computed Tomography (CT) analysis remains constrained by a persistent mismatch between optimization objectives and clinical rigor. Current Reinforcement Learning (RL) paradigms still rely on lexical proxy signals that induce Evaluation Hallucinations, where models optimize linguistic fluency rather than factual clinical correctness, leading to diagnostically critical errors. To bridge this gap, we introduce the Clinical Abnormality Benchmarking Substrate (CABS), a structured system that decomposes radiology reports into verifiable clinical semantic units. Using CABS, we identify a Mechanistic Divergence in standard RL, where surface-similarity rewards drive policy gradients to bypass medical facts. We therefore propose Trajectory-Integral Feedback GRPO (TIF-GRPO), a novel framework integrating control-theoretic principles into policy optimization. By formulating clinical reasoning as a pseudo-temporal trajectory for anomaly discovery, TIF-GRPO regulates anatomy-aware rewards via an integral feedback loop that penalizes persistent omissions as cumulative state errors and suppresses hallucinations as excessive control effort. Experiments on 3D CT benchmarks demonstrate that our approach significantly enhances abnormality detection and clinical faithfulness, establishing a new paradigm for fine-grained regulation in medical VLMs. Our project is available at https://github.com/ZJU4HealthCare/TIF-GRPO.

Federated parameter-efficient fine-tuning (PEFT) enables customizing large language models on private data, yet it is vulnerable to backdoor poisoning—especially when privacy constraints prevent inspection of per-client real-valued updates. We exploit the intuition that poisoning leaves a similar backdoor imprint in which adapter coordinates become salient, so overlap in salient-index supports remains informative even without values. We introduce IndexGuard, an unsupervised index-only vetting primitive in which clients send only Top-$K$ salient update indices and the server operates on the induced overlap geometry, clustering clients and filtering cohesion-outlier groups before aggregation. We analyze support stability under bounded rescaling and separability under shared-trigger poisoning under non-IID drift. Across attacks, backbones, and PEFT variants, IndexGuard provides end-to-end mitigation, preserving clean accuracy while achieving performance comparable to centralized methods.

While prompt engineering is instrumental in maximizing the capabilities of Large Language Models (LLMs) during inference, the role of prompts during training remains critically underexplored. Prevailing fine-tuning paradigms typically treat training prompts as mere surface forms, assuming that semantically equivalent instructions yield identical learning outcomes. However, we reveal that this equivalence is deceptive: while paraphrased prompts often lead to comparable in-task performance, they induce drastically different cross-task impacts regarding catastrophic forgetting and generalization. Crucially, these impacts are positively correlated across tasks, indicating the existence of superior prompts that consistently yield better performance. Furthermore, we discover that these superior prompts can be robustly identified by task loss prior to learning. Leveraging these insights, we introduce State-Adaptive Prompt Optimization (SAPO), a lightweight yet effective training strategy that shifts task formulation from a static input to a dynamic, state-adaptive variable. Comprehensive experiments on diverse benchmarks confirm its effectiveness, which significantly mitigates forgetting while improving generalization, achieving substantial performance gains over state-of-the-art methods. These results provide insights into how training prompts shape learning dynamics and offer a practical recipe for robust fine-tuning. Our code is available at https://github.com/Eric8932/SAPO.

In machine unlearning, $(\varepsilon,\delta)-$unlearning is a popular framework that provides formal guarantees on the effectiveness of the removal of a subset of training data, the \emph{forget set}, from a trained model. For strongly convex objectives, existing first-order methods achieve $(\varepsilon,\delta)-$unlearning, but they only use the forget set to calibrate injected noise, never as a direct optimization signal. In contrast, efficient empirical heuristics often exploit the forget samples (e.g., via gradient ascent) but come with no formal unlearning guarantees. We bridge this gap by presenting the Variance-Reduced Unlearning (*VRU*) algorithm. To the best of our knowledge, *VRU* is the first first-order algorithm that directly includes forget set gradients in its update rule, while provably satisfying $(\varepsilon,\delta)-$unlearning. We establish the convergence of *VRU* and show that incorporating the forget set yields strictly improved rates, *i.e.*, a better dependence on the achieved error compared to existing first-order $(\varepsilon,\delta)-$unlearning methods. Moreover, we prove that, in a low-error regime *VRU* asymptotically outperforms any first-order methods that ignores the forget set. Experiments corroborate our theory, showing consistent gains over both state-of-the-art certified unlearning methods and over empirical baselines that explicitly leverage the forget set.

Transformer-based diffusion models have demonstrated remarkable performance at generating high-quality samples. However, our theoretical understanding of the reasons for this success remains limited. For instance, existing models are typically trained by minimizing a denoising objective, which is equivalent to fitting the score function of the training data. However, we do not know why transformer-based models can match the score function for denoising, or why gradient-based methods converge to the optimal denoising model despite the non-convex loss landscape. To the best of our knowledge, this paper provides the first convergence analysis for training transformer-based diffusion models. More specifically, we consider the population Denoising Diffusion Probabilistic Model (DDPM) objective for denoising data that follow a \textit{multi-token Gaussian mixture} distribution. We theoretically quantify the required number of tokens per data point and training iterations for the global convergence towards the Bayes optimal risk of the denoising objective, thereby achieving a desired score matching error. A deeper investigation reveals that the self-attention module of the trained transformer implements a \emph{mean denoising} mechanism that enables the trained model to approximate the oracle Minimum Mean Squared Error (MMSE) estimator of the injected noise in the diffusion steps. Numerical experiments validate these findings.

Generative modeling within constrained sets is essential for scientific and engineering applications involving physical, geometric, or safety requirements (e.g., molecular generation, robotics). We present a unified framework for constrained diffusion models on generic nonconvex feasible sets $\Sigma$ that simultaneously enforces equality and inequality constraints throughout the diffusion process. Our framework incorporates both overdamped and underdamped dynamics for forward and backward sampling. A key algorithmic innovation is a computationally efficient landing mechanism that replaces costly and often ill-defined projections onto $\Sigma$, ensuring feasibility without iterative Newton solves or projection failures. By leveraging underdamped dynamics, we accelerate mixing toward the prior distribution, effectively alleviating the high simulation costs typically associated with constrained diffusion. Empirically, this approach reduces function evaluations and memory usage during both training and inference while preserving sample quality. On benchmarks featuring equality and mixed constraints, our method achieves comparable sample quality to state-of-the-art baselines while significantly reducing computational cost, providing a practical and scalable solution for diffusion on nonconvex feasible sets.

Unified Foundation Models (UFMs), which support interleaved multimodal generation and understanding, have been proposed as a promising paradigm for reasoning about dynamic world states, yet it remains unclear whether the visual content they generate functions as grounded evidence for subsequent reasoning or merely as auxiliary output. Existing benchmarks largely evaluate generation and understanding as separate capabilities and do not test their functional dependence during reasoning. We introduce UFO, a benchmark designed to evaluate whether UFMs generate and use image and text cues as evidence for compositional multimodal reasoning. UFO spans three state-transition regimes, state determination, state reconstruction, and state augmentation, which correspond to progressively smaller transformations of the underlying world state. Our analysis reveals a significant modality gap, as models often achieve high prediction accuracy even when the generated visual cues exert limited influence on their decisions, indicating weakened evidential coupling and a reliance on textual shortcuts rather than robust cross-modal grounding.

Aligning Large Language Models (LLMs) with specific personas typically relies on Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF); however, these methods are resource-intensive, requiring expensive data collection and distinct model training for each target personality. In this work, we propose a parameter-efficient framework for continuous, multi-dimensional personality control via inference-time activation steering. Our approach addresses the challenge of combining multiple interventions by iteratively retraining probes on the residual stream modified by prior traits, ensuring compatibility. Once established, these steering vectors function as modular, reusable primitives; users can instantly synthesize novel, complex personality profiles by simply adjusting steering coefficients (α) without any additional training. To support this, we introduce an automated pipeline that identifies optimal intervention layers via activation separation analysis and calibrates coefficients via hyperparameter optimization to maximize alignment while constraining perplexity. Empirical evaluations validate individual trait shifts using an LLM-as-a-judge framework and demonstrate, via the Big Five inventory, that our method effectively modulates the model's holistic personality profile without updating base model parameters.

The online construction of vectorized high-definition (HD) maps is a cornerstone of modern autonomous driving systems. State-of-the-art approaches, particularly those based on the DETR framework, formulate this as an instance detection problem. However, their reliance on independent, learnable object queries results in a predominantly local query perspective, neglecting the inherent global representation within HD maps. In this work, we propose \textbf{MapGR} (\textbf{G}lobal \textbf{R}epresentation learning for HD \textbf{Map} construction), an architecture designed to learn and utilize global representations from queries. Our method introduces two synergistic modules: a Global Representation Learning (GRL) module, which encourages the distribution of all queries to better align with the global map through a carefully designed holistic segmentation task, and a Global Representation Guidance (GRG) module, which endows each individual query with explicit, global-level contextual information to facilitate its optimization. Evaluations on the nuScenes and Argoverse2 datasets validate the efficacy of our approach, demonstrating substantial improvements in mean Average Precision (mAP) compared to leading baselines. Code is available at \href{https://github.com/skyshoumeng/MapGR}{\texttt{github.com/skyshoumeng/MapGR}}.

The rapid growth of deep neural networks (DNNs) has intensified the demand for efficient hardware acceleration under quantization. While prior research has successfully reduced weight and activation precision, partial sums generated during accumulation often retain high precision, resulting in significant energy overhead. In this work, we analyze psum distributions in tiled architectures and reveal that within-tile outliers are input-dependent. We propose PsumQuant, a post-training, input-aware quantization that predicts psum scales on-the-fly. By leveraging the crest factor of input activations, our learnable scale predictor effectively bounds the psum bit-width while handling the extreme outliers in DNNs. Experimental results on a $128 \times 128$ systolic array demonstrate that PsumQuant compresses psum precision down to 8-bit within only a 1\% accuracy drop on ResNet-18 and a marginal 0.04 perplexity increase on Llama-3.1. Furthermore, bit-width reduction with PsumQuant results in a 45\% reduction in total energy with minimal accuracy loss, demonstrating that PsumQuant provides a highly efficient solution for actual NPU architectures.

Mixture-of-Experts (MoE) architectures are evolving towards finer granularity to improve parameter efficiency. However, existing MoE designs face an inherent trade-off between the granularity of expert specialization and hardware execution efficiency. In this paper, we propose OmniMoE, a system-algorithm co-designed MoE framework that pushes granularity to the extreme with vector-level Atomic Experts, orchestrating their routing and execution at scale within a single MoE layer, while retaining a shared dense MLP for general-purpose processing. While this atomic design maximizes capacity, it poses severe challenges for routing complexity and memory access. To address these, OmniMoE adopts a system-algorithm co-design: (i) a Cartesian Product Router that decomposes the massive index space to reduce routing complexity from $O(N)$ to $O(\sqrt{N})$; and (ii) Expert-Centric Scheduling that inverts the execution order to turn scattered, memory-bound lookups into efficient dense matrix operations. Validated on seven benchmarks, OmniMoE (with 1.7B active parameters) achieves 50.9\% zero-shot accuracy across seven benchmarks, outperforming coarse-grained (e.g., DeepSeekMoE) and fine-grained (e.g., PEER) baselines. Crucially, OmniMoE reduces inference latency from 73ms to 6.7ms (a 10.9$\times$ speedup) compared to PEER, demonstrating that massive-scale fine-grained MoE can be fast and accurate.

Personality traits have long been studied as predictors of human behavior. Recent advances in Large Language Models (LLMs) suggest similar patterns may emerge in artificial systems, with advanced LLMs displaying consistent behavioral tendencies resembling human traits like agreeableness and self-regulation. Understanding these patterns is crucial, yet prior work primarily relied on simplified self-reports and heuristic prompting, with little behavioral validation. In this study, we systematically characterize LLM personality across three dimensions: (1) the dynamic emergence and evolution of trait profiles throughout training stages; (2) the predictive validity of self-reported traits in behavioral tasks; and (3) the impact of targeted interventions, such as persona injection, on both self-reports and behavior. Our findings reveal that instructional alignment (e.g., RLHF, instruction tuning) significantly stabilizes trait expression and strengthens trait correlations in ways that mirror human data. However, these self-reported traits do not reliably predict behavior, and observed associations often diverge from human patterns. While persona injection successfully steers self-reports in the intended direction, it exerts little or inconsistent effect on actual behavior. By distinguishing surface-level trait expression from behavioral consistency, our findings challenge assumptions about LLM personality and underscore the need for deeper evaluation in alignment and interpretability.

Transformer architectures have established strong baselines in time series forecasting, yet they typically rely on positional encodings that assume uniform, index-based temporal progression. However, real-world systems, from shifting financial cycles to elastic biological rhythms, frequently exhibit ``time-warped'' dynamics where the effective flow of time decouples from the sampling index. In this work, we first formalize this misalignment and prove that rotary position embedding (RoPE) is mathematically incapable of representing non-affine temporal warping. To address this, we propose Symplectic Positional Embeddings (SyPE), a learnable encoding framework derived from Hamiltonian mechanics. SyPE strictly generalizes RoPE by extending the rotation group $\mathrm{SO}(2)$ to the symplectic group $\mathrm{Sp}(2,\mathbb{R})$, modulated by a novel input-dependent adaptive warp module. By allowing the attention mechanism to adaptively dilate or contract temporal coordinates end-to-end, our approach captures locally varying periodicities without requiring pre-defined warping functions. We implement this mechanism in StretchTime, a multivariate forecasting architecture that achieves state-of-the-art performance on standard benchmarks, demonstrating superior robustness on datasets exhibiting non-stationary temporal dynamics.

We present $\textit{Video-in-the-Loop}$ (ViTL), a two-stage long-video QA framework that preserves a fixed token budget by first $\textit{localizing}$ question-relevant interval(s) with a low-fps skim and then $\textit{answering}$ via span-aware reallocation of visual tokens at higher effective frame rate, emitting an interleaved output with both spans and the final option for direct attribution. We also introduce $\textit{VGrounding-QA}$, which converts description based event graphs into $\textit{span-grounded}$ multiple-choice QA by pairing each question with $\textit{ground-truth}$ time span(s) and related reasoning. ViTL is trained end-to-end with an interleaved group-relative objective that couples temporal IoU for localization with answer correctness, allowing credit to flow from answers back to spans without increasing compute. Under fixed token budgets, ViTL attains up to 8.6\% with 50\% less frame input on long-video QA and temporal grounding (e.g., Charades-STA, ActivityNet-Captions) and ablations show that span-aware token reallocation consistently surpasses uniform sampling. Together, $\textit{VGrounding-QA}$ and ViTL provide an interpretable, compute-efficient recipe for scalable long-video QA.

Emerging agentic marketplaces provide the economic infrastructure for matching and coordinating the large amounts of AI agents used in agentic swarms. Unlike human workers, AI agents can operate on multiple jobs simultaneously, acquire skills rapidly, and labor without wage floors. These differences introduce a new segment of AI labor markets, where AI agents interact with each other at a much higher frequency than human markets. Yet we lack frameworks to understand how such markets behave in light of economic forces that shape labor markets, such as adverse selection and reputation dynamics. To explore this, we introduce AI-Work, a tractable, simulated gig economy where Large Language Model (LLM) agents compete for jobs, develop skills, and adapt their strategies under uncertainty and competitive pressure. Our experiments examine three domains of capabilities that successful agents possess: metacognition (accurate self-assessment of skills), competitive awareness (modeling rivals and market dynamics), and long-horizon strategic planning. Agents with these capabilities consistently achieve higher profits, reputations, and market share than competing agents. Through AI-Work, we hope to provide a foundation to explore the microeconomic properties of AI-only labour markets, and a conceptual framework to study the strategic reasoning capabilities of participating AI agents.

Planning with a learned model remains a key challenge in model-based reinforcement learning (RL). In decision-time planning, state representations are critical as they must support local cost computation while preserving long-horizon structure. In this paper, we show that the Laplacian representation provides an effective latent space for planning by capturing state-space distances at multiple time scales. This representation preserves meaningful distances and naturally decomposes long-horizon problems into subgoals, also mitigating the compounding errors that arise over long prediction horizons. Building on these properties, we introduce ALPS, a hierarchical planning algorithm, and demonstrate that it outperforms commonly used baselines on a selection of offline goal-conditioned RL tasks from OGBench, a benchmark previously dominated by model-free methods.

Instruction-based image editing (IIE) is a vital tool for iterative content creation, enabling multi-round interactions that refine visual details while preserving cross-round consistency. However, this workflow is constrained by the compute-bound nature of Diffusion Transformers (DiTs): because DiTs process tokens uniformly, they waste substantial computation on regions untouched by the instruction. We investigate the Round--Step--Layer hierarchy of DiT-based editing and identify a phenomenon we term Delayed Latent Emergence (DLE). Although pronounced latent changes emerge only in the late denoising stages, deep-layer activations within transformer blocks at the very first sampling step already diverge markedly in edited regions. Building on this insight, we propose CacheEdit, a training-free framework centered on an Adaptive Activation Cache (Acache) that exploits early-step sensitivity to detect invariant tokens and reuse their cached activations across subsequent sampling steps, thereby bypassing redundant computation. Experiments on FLUX.1 Kontext and Qwen-Image-Edit show that CacheEdit achieves up to $2.5\times$ end-to-end acceleration. Moreover, by isolating and reusing static features, CacheEdit mitigates stochastic drift and improves instruction-following and structural consistency over full-recomputation baselines.