ICML 2026 论文列表

Diffusion planners are a strong approach for offline reinforcement learning, but they can fail when value-guided selection favours trajectories that score well yet are locally inconsistent with the environment dynamics, resulting in brittle execution. We propose Self-supervised Action Gating with Energies (SAGE), an inference-time re-ranking method that penalises dynamically inconsistent plans using a latent consistency signal. SAGE trains a Joint-Embedding Predictive Architecture (JEPA) encoder on offline state sequences and an action-conditioned latent predictor for short horizon transitions. At test time, SAGE assigns each sampled candidate an energy given by its latent prediction error and combines this feasibility score with value estimates to select actions. SAGE can integrates into existing diffusion planning pipelines that can sample trajectories and select actions via value scoring; it requires no environment rollouts and no policy re-training. Across locomotion, navigation, and manipulation benchmarks, SAGE improves the performance and robustness of diffusion planners.

Imitation learning with diffusion models has advanced robotic control by capturing the multi-modal action distributions. However, existing methods typically treat observations only as high-level conditions to the denoising network, rather than integrating them into the stochastic dynamics of the diffusion process itself. As a result, the sampling is forced to begin from random noise, weakening the coupling between perception and control and often yielding suboptimal performance. We propose BridgePolicy, a generative visuomotor policy that directly integrates observations into the stochastic dynamics via a diffusion-bridge formulation. By constructing an observation-informed trajectory, BridgePolicy enables sampling to start from a rich and informative prior rather than random noise, substantially improving precision and reliability in control. A key difficulty is that diffusion bridge normally connects distributions of matched dimensionality, while robotic observations are heterogeneous and not naturally aligned with actions. To overcome this, we introduce a semantic aligner to unify the visual and state inputs and align the observations with action representations, making diffusion bridge applicable to heterogeneous robot data. Extensive experiments across 52 simulation tasks on three benchmarks and 5 real-world tasks demonstrate that BridgePolicy consistently outperforms state-of-the-art generative policies. Our code is available at https://github.com/jianghcsr/BridgePolicy.

In this paper, we study cooperative multi-agent reinforcement learning (MARL) where the joint reward exhibits submodularity, which is a natural property capturing diminishing marginal returns when adding agents to a team. Unlike standard MARL with additive rewards, submodular rewards model realistic scenarios where agent contributions overlap (e.g., multi-drone surveillance, collaborative exploration). We provide the first formal framework for this setting and develop algorithms with provable guarantees on sample efficiency and regret bound. For known dynamics, our greedy policy optimization achieves a $1/2$-approximation with polynomial complexity in the number of agents $K$, overcoming the exponential curse of dimensionality inherent in joint policy optimization. For unknown dynamics, we propose a UCB-based learning algorithm achieving a $1/2$-regret of $O(H^2KS\sqrt{AT})$ over $T$ episodes.

Long-duration streaming video understanding is fundamental for future AI agents, yet remains limited by ineffective long-term memory. We introduce video-SALMONN S, a memory-enhanced streaming audio-visual large language model that processes over 3-hour videos at $1$ FPS and $360$p resolution, outperforming strong non-streaming models under the same memory budget. In addition to token merging or downsampling, video-SALMONN S is the first to employ test-time training (TTT) as a streaming memory mechanism for video understanding. TTT continuously transforms short-term multimodal representations into long-term memory embedded in model parameters. To improve long-range dependency modeling and memory capacity, we propose (i) a TTT$_\text{MEM}$ layer with an additional long-span prediction objective, (ii) a two-stage training scheme, and (iii) a modality-aware memory reader. We further introduce the episodic learning from video memory (ELViM) benchmark, simulating agent-like scenarios where models must learn from videos observed hours earlier. video-SALMONN S consistently outperforms both streaming and non-streaming baselines by 3-7\% on long video benchmarks. Notably, video-SALMONN S achieves a $15\%$ absolute accuracy improvement over strong non-streaming models on ELViM, demonstrating strong learning abilities from video memory.

We consider the computational problem of compacting Tsetlin Machine classifiers by reducing the number of propositional clauses while preserving predictive accuracy. TMs trained with limited clause capacity often perform poorly because stochastic optimization cannot reliably find the few precise clauses needed in a vast configuration space. High-quality compact subsets also exist in the case of larger Tsetlin Machines. The difficulty here lies in extracting them. Local pruning heuristics can fail badly on TMs because clauses interact through Boolean logic: a clause may appear unimportant in isolation yet becomes critical when others are removed. We formalize compression as the Minimum Discriminating Clause Set (MDCS) problem, which asks to find a smallest subset of clauses that preserves the trained model's discrimination of training samples. We show that MDCS is NP-hard. We solve MDCS using weighted partial Maximum Satisfiability (MaxSAT). A partition-and-merge strategy allows us to scale to 100,000 samples. Across 13 datasets, the compressed model preserves the 200-clause teacher's accuracy within a few percentage points while using a median of only 16 clauses, and outperforms a matched-capacity TM trained from scratch on every dataset where direct training has room to improve, by up to 45 percentage points.

$\textit{De novo}$ peptide sequencing is a pivotal technique that directly reconstructs amino acid sequences from tandem mass spectrometry (MS/MS) data; it enables the identification of novel proteins and variants absent from reference databases. Previous methods are typically based on autoregressive (AR) decoding or one-shot generation. The AR-based methods conflict with the bidirectional and globally constrained nature of MS/MS evidence and inevitably accumulate errors, while one-shot generation does not explicitly enforce physical constraints, failing to produce chemically valid and reliable peptides in a single pass. Accurate sequencing necessitates reasoning over the entire peptide simultaneously, enabling iterative self-correction under global constraints. To this end, we introduce $\textbf{PhysNovo}$, a novel paradigm that harnesses discrete diffusion to enable simultaneous global reasoning and iterative refinement. Specifically, PhysNovo reformulates sequencing as a $\textbf{phys}$ically mass-constrained reasoning process by embedding a knapsack-based feasibility kernel to enforce exact precursor mass consistency. By conditioning the diffusion process on global spectral context, PhysNovo supports abductive reasoning where bidirectional evidence is exploited to iteratively resolve local inconsistencies and ensure physically valid predictions. PhysNovo achieves state-of-the-art performance, exceeding baselines by over 2% in precision, with larger gains on out-of-distribution data. The source code is publicly available at https://github.com/WanyuGroup/ICML2026_PhysNovo.

Large Language Models (LLMs) often provide chain-of-thought (CoT) reasoning traces that appear plausible, but may hide internal biases. We call these *unverbalized biases*. Monitoring models via their stated reasoning is therefore unreliable, and existing bias evaluations typically require predefined categories and hand-crafted datasets. In this work, we introduce a fully automated, black-box pipeline for detecting task-specific unverbalized biases. Given a task dataset, the pipeline uses LLM autoraters to generate candidate bias concepts. It then tests each concept on progressively larger input samples by generating positive and negative variations, and applies statistical techniques for multiple testing and early stopping. A concept is flagged as an unverbalized bias if it yields statistically significant performance differences while not being cited as justification in the model's CoTs. We evaluate our pipeline across seven LLMs on three decision tasks (hiring, loan approval, and university admissions). Our technique automatically discovers previously unknown biases in these models (e.g., Spanish fluency, English proficiency, writing formality). In the same run, the pipeline also validates biases that were manually identified by prior work (gender, race, religion, ethnicity). More broadly, our proposed approach provides a practical, scalable path to automatic, more efficient, and broader task-specific unverbalized bias discovery.

Code complexity metrics such as cyclomatic complexity have long been used to assess software quality and maintainability. With the rapid advancement of large language models (LLMs) on coding tasks, an important yet underexplored question arises: do traditional complexity metrics meaningfully characterize the coding difficulty that LLMs perceive? In this work, we empirically demonstrate that classical complexity metrics exhibit no consistent correlation with LLM performance, revealing a fundamental mismatch with model-perceived difficulty. To address this gap, we propose LM-CC, a novel code complexity metric tailored for LLMs, grounded in the hypothesis that model-perceived code difficulty is fundamentally driven by semantic nonlinearity. LM-CC quantifies complexity through an entropy-guided semantic compositional hierarchy, capturing the cumulative uncertainty encountered by LLMs during code understanding. Our experimental results demonstrate that LM-CC exhibits strong and consistent partial correlations with LLM performance, while semantics-preserving reductions in LM-CC consistently lead to improved downstream task performance. The source code is available at: https://github.com/xchen121/lm-cc.

Aligning large language models to human preferences is inherently multidimensional, yet most pipelines collapse heterogeneous signals into a single objective. We seek to answer what it would take to simultaneously align a model across various domains spanning those with: verifiable rewards, non-verifiable subjective preferences, and complex interactive scenarios. Such multi-objective alignment setups are often plagued by individual objectives being at odds with each other, resulting in inefficient training and limited user control during inference. To address these issues, we propose **M**ulti-**A**ction-**H**ead **AL**ignment with PRM-guided Dec**O**ding (**MAHALO**), a unified framework that standardizes PRM training across verifiable and non-verifiable settings for step-level supervision, performs vectorized multi-objective alignment with Multi-Action-Head DPO, and enables controllable inference through objective-specific weighting and PRM-guided decoding. Experiments across math reasoning, human values alignment, and multi-turn tutoring show that MAHALO jointly improves multiple objectives simultaneously with limited interference, while remaining generalizable and adaptable across domains and offering flexible user control at inference time. Our code is available at: https://github.com/pearls-lab/multiobj-align.

Online Multi-Agent Reinforcement Learning (MARL) is a prominent framework for efficient agent coordination. Crucially, enhancing policy expressiveness is pivotal for achieving superior performance. Diffusion-based generative models are well-positioned to meet this demand, having demonstrated remarkable expressiveness and multimodal representation in image generation and offline settings. Yet, their potential in online MARL remains largely under-explored. A major obstacle is that the intractable likelihoods of diffusion models impede entropy-based exploration and coordination. To tackle this challenge, we propose among the first Online off-policy MARL framework using Diffusion policies (**OMAD**) to orchestrate coordination. Our key innovation is a relaxed policy objective that maximizes scaled joint entropy, facilitating effective exploration without relying on tractable likelihood. Complementing this, within the centralized training with decentralized execution (CTDE) paradigm, we employ a joint distributional value function to optimize decentralized diffusion policies. It leverages tractable entropy-augmented targets to guide the simultaneous updates of diffusion policies, thereby ensuring stable coordination. Extensive evaluations on MPE and MAMuJoCo establish our method as the new state-of-the-art across $10$ diverse tasks, demonstrating a remarkable $2.5\times$ to $5\times$ improvement in sample efficiency.

Feature-importance methods show promise for transforming machine learning (ML) models from predictive engines into tools for scientific discovery. However, expressive models can be unstable due to data sampling and algorithmic stochasticity, leading to inaccurate variable importance estimates, undermining their utility in critical biomedical applications. While ensembling offers a remedy, the choice between explaining a single ensemble model or aggregating individual model explanations is non-trivial due to the non-linearity of importance measures, and remains largely understudied. Our theoretical analysis, developed under assumptions accommodating complex state-of-the-art ML models, reveals that this choice is governed by a trade-off involving the model's excess risk. In contrast to prior literature, we show that ensembling at the model level provides more accurate variable-importance estimates, particularly for expressive models, by reducing this leading error term. We validate these findings on classical benchmarks and a large-scale proteomic study from the UK Biobank.

In RL post-training of LLM agents, calls to external tools take several seconds or even minutes, leaving allocated GPUs idle and inflating post-training time and cost. While many tool invocations repeat across parallel rollouts and could in principle be cached, naively caching their outputs for reuse is incorrect since tool outputs depend on the environment state induced by prior agent interactions. We present TVCACHE, a stateful tool-value cache for LLM agent post-training. TVCACHE maintains a tree of observed tool-call sequences and performs longest-prefix matching for cache lookups: a hit occurs only when the agent’s full tool history matches a previously executed sequence, guaranteeing identical environment state. On three diverse workloads—terminal-based tasks, SQL generation, and video understanding—TVCACHE achieves cache hit rates of up to 70% and reduces median tool call execution time by up to 6.9×, with no degradation in post-training reward accumulation.

Existing 3D Visual Question Answering (3D-VQA) methods rely on generative outputs that can be ambiguous in decision-making settings. We introduce 3D Scene Assertion Verification, a task that verifies natural language assertions in 3D scenes with strict binary judgments. We present 3DSAV, a large-scale diagnostic benchmark with 22.5k samples across six semantic types. To address this task, we propose DualLPSS, which uses dual-stage subspace routing for type-aware cross-modal fusion and scene-guided assertion focusing. Experiments show that DualLPSS achieves state-of-the-art performance on 3DSAV and handles complex logical assertions better than existing 3D-VQA baselines.

Direct Alignment Algorithms (DAAs) simplify LLM alignment by directly optimizing policies, bypassing reward modeling and RL. While DAAs differ in their use of SFT (one-stage vs. two-stage) and the scalar score they optimize (likelihood vs. odds ratios), the key performance drivers remain underexplored. We present a systematic comparison and analyze a previously overlooked axis - the ranking objective (pairwise vs. pointwise). To isolate this factor, we propose a unified training framework across DAAs by (i) converting one-stage methods (ORPO, ASFT) into a two-stage pipeline with an explicit SFT phase and (ii) introducing a $\beta$ parameter that places all methods in the same hyperparameter space and improves the quality of odds-ratio DAAs (ORPO, ASFT). Under this setup, the ranking objective emerges as the primary determinant of alignment quality, whereas the particular scalar score (policy–reference ratio vs. odds ratio) is secondary. We corroborate this on instruction-following tasks and further confirm it on math-reasoning benchmarks across model scales. Evidence suggests that this stems from how these objectives interact with prompt-specific biases, supported both by strictly controlled experiments and by observations on real data. Our findings underscore the need for nuanced evaluations in DAA research to avoid oversimplified claims of superiority.

Modern database workloads are highly predictable: query streams are dominated by recurring jobs and templates, even when their arrival order is not known in advance. This motivates a learning-augmented view of online differentially private (DP) analytics: can algorithms utilize predictions about *which* queries will occur to improve utility under a single global privacy budget, while remaining robust when predictions are wrong? We study online DP query answering, where a curator must answer a stream $Q$ of $S$ linear queries arriving in uniformly random order under privacy budget $(\epsilon,\delta)$. We present *LAPRAS*, which assumes access to an oracle that outputs a prediction set of queries likely to appear in the stream and uses it to guide privacy spending. LAPRAS answers predicted queries using the offline-optimal Matrix Mechanism and answers the remaining queries online from a residual budget. To pace spending across an unknown number of unpredicted queries, we introduce *Smooth Allocation*, which forms an unbiased stopping-time estimate $\widehat{B}$ from the first $T=\Theta(\log^2 S)$ unpredicted queries and continuously recalibrates per-query expenditure. Empirically, over two real datasets, we validate the intended consistency--robustness trade-off: LAPRAS achieves near-offline utility under high overlap and degrades gracefully to baseline-level performance when overlap is low.

Inference-time scaling offers a flexible way to align visual generative models with downstream objectives without updating model parameters. In modern image generation models, a natural way to do this is to optimize the random noise from which generation starts. However, searching in this high-dimensional noise space is highly inefficient, because many directions have little effect on the final image. We trace this inefficiency to a spectral bias in generative dynamics: model sensitivity to initial perturbations decays rapidly as frequency increases. Based on this insight, we propose Spectral Evolution Search (SES), a plug-and-play framework for initial noise optimization that performs gradient-free evolutionary search in a low-frequency subspace. Theoretically, we derive the Spectral Scaling Prediction from perturbation propagation dynamics, which explains the frequency-dependent impact of perturbations. Extensive experiments across diverse settings show that SES substantially improves the trade-off between generation quality and computational cost, consistently outperforming strong baselines under the same compute budget.

Humans are born with vision-based 4D spatial-temporal intelligence, which enables us to perceive and reason about the evolution of 3D space over time from purely visual inputs. Despite its importance, this capability remains a significant bottleneck for current multimodal large language models (MLLMs). To tackle this challenge, we introduce MLLM-4D, a comprehensive framework designed to bridge the gaps in training data curation and model post-training for spatiotemporal understanding and reasoning. On the data front, we develop a cost-efficient data curation pipeline that repurposes existing stereo video datasets into high-quality 4D spatiotemporal instructional data. This results in the MLLM4D-2M and MLLM4D-R1-30k datasets for Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT), alongside MLLM4D-Bench for comprehensive evaluation. Regarding model training, our post-training strategy establishes a foundational 4D understanding via SFT and further catalyzes 4D reasoning capabilities by employing Group Relative Policy Optimization (GRPO) with specialized Spatiotemporal Chain of Thought (ST-CoT) prompting and Spatiotemporal reward functions (ST-reward) without involving the modification of architecture. Extensive experiments demonstrate that MLLM-4D achieves state-of-the-art spatial-temporal understanding and reasoning capabilities from purely 2D RGB inputs. Project page: https://github.com/GVCLab/MLLM-4D.

We study fully dynamic non-monotone submodular maximization under a cardinality constraint $k$. Prior work achieved approximation guarantees of $(0.125-\epsilon)$ using $\tilde{O}(\epsilon^{-1}k^2)$ oracle queries per update (NeurIPS'20) and $0.171$ using $\tilde{O}(\epsilon^{-3}k^4)$ oracle queries per update (NeurIPS'25). In this work, we present a dynamic algorithm that achieves a $0.262$-approximation with worst-case expected update time $O(\epsilon^{-3}k\log(k)\log(\epsilon^{-1}k) + \epsilon^{-2}k^2\log(k))$, where $0 < \epsilon \leq 1$ is the error parameter. We also develop another dynamic algorithm with update time bounded by $\mathrm{poly}(\epsilon^{-1},k)$ that achieves a $0.277$-approximation guarantee.

LLM-based agents execute real-world workflows via tools. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent-misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce **STING** (*Sequential Testing of Illicit N-step Goal execution*), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do *not* consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual. Our code is available at https://github.com/epfl-nlp/helpful-to-a-fault.