ICLR 2026 Papers

We study equilibrium finding in polymatrix games under differential privacy constraints. Prior work in this area fails to achieve both high-accuracy equilibria and a low privacy budget. To better understand the fundamental limitations of differential privacy in games, we show hardness results establishing that no algorithm can simultaneously obtain high accuracy and a vanishing privacy budget as the number of players tends to infinity. This impossibility holds in two regimes: (i) We seek to establish equilibrium approximation guarantees in terms of Euclidean \emph{distance} to the equilibrium set, and (ii) The adversary has access to all communication channels. We then consider the more realistic setting in which the adversary can access only a bounded number of channels and propose a new distributed algorithm that: recovers strategies with simultaneously vanishing \emph{Nash gap} (in expected utility, also referred to as \emph{exploitability}) and \emph{privacy budget} as the number of players increases. Our approach leverages structural properties of polymatrix games. To our knowledge, this is the first paper that can achieve this in equilibrium computation. Finally, we also provide numerical results to justify our algorithm.

Knowledge distillation has emerged as a pivotal technique for transferring knowledge from stronger large language models (LLMs) to smaller, more efficient models. However, traditional distillation approaches face challenges related to knowledge conflicts and high resource demands, particularly when leveraging multiple teacher models. In this paper, we introduce the concept of **Knowledge Purification**, which consolidates the rationales from multiple teacher LLMs into a single rationale, thereby mitigating conflicts and enhancing efficiency. To investigate the effectiveness of knowledge purification, we further propose five purification methods from various perspectives. Our experiments demonstrate that these methods not only improve the performance of the distilled model but also effectively alleviate knowledge conflicts. Moreover, router-based methods exhibit robust generalization capabilities, underscoring the potential of innovative purification techniques in optimizing multi-teacher distillation and facilitating the practical deployment of powerful yet lightweight models.

Recent advances in video generation have opened new avenues for macroscopic simulation of complex dynamic systems, but their application to microscopic phenomena remains largely unexplored. Microscale simulation holds great promise for biomedical applications such as drug discovery, organ-on-chip systems, and disease mechanism studies, while also showing potential in education and interactive visualization. In this work, we introduce **MicroWorldBench**, a multi-level rubric-based benchmark for microscale simulation tasks. MicroWorldBench enables systematic, rubric-based evaluation through 459 unique expert-annotated criteria spanning multiple microscale simulation task (e.g., organ-level processes, cellular dynamics, and subcellular molecular interactions) and evaluation dimensions (e.g., scientific fidelity, visual quality, instruction following). MicroWorldBench reveals that current SOTA video generation models fail in microscale simulation, showing violations of physical laws, temporal inconsistency, and misalignment with expert criteria. To address these limitations, we construct **MicroSim-10K**, a high-quality, expert-verified simulation dataset built with expert verification. Leveraging this dataset, we train **MicroVerse**, a video generation model tailored for microscale simulation. MicroVerse can accurately reproduce complex microscale phenomena. Our work first introduce the concept of **Micro-World Simulation** and present a **proof of concept**, paving the way for applications in biology, education, and scientific visualization. Our work demonstrates the potential of educational microscale simulations of biological mechanisms.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks, but optimizing LLM-based agentic systems remains challenging due to the vast search space of agent configurations, prompting strategies, and communication patterns. Existing approaches often rely on heuristic-based tuning or exhaustive evaluation, which can be computationally expensive and suboptimal. This paper proposes **Agentic Predictor**, a lightweight predictor for efficient agentic workflow evaluation. Agentic Predictor is equipped with a *multi-view workflow encoding* technique that leverages multi-view representation learning of agentic systems by incorporating code architecture, textual prompts, and interaction graph features. To achieve high predictive accuracy while significantly reducing the number of required workflow evaluations for training a predictor, Agentic Predictor employs *cross-domain unsupervised pretraining*. By learning to approximate task success rates, Agentic Predictor enables fast and accurate selection of optimal agentic workflow configurations for a given task, significantly reducing the need for expensive trial-and-error evaluations. Experiments on a carefully curated benchmark spanning three domains show that our predictor outperforms state-of-the-art methods in both predictive accuracy and workflow utility, highlighting the potential of performance predictors in streamlining the design of LLM-based agentic workflows.

We develop a novel method for personalized off-policy learning in scenarios with unobserved confounding. Thereby, we address a key limitation of standard policy learning: standard policy learning assumes unconfoundedness, meaning that no unobserved factors influence both treatment assignment and outcomes. However, this assumption is often violated, because of which standard policy learning produces biased estimates and thus leads to policies that can be harmful. To address this limitation, we employ causal sensitivity analysis and derive a semi-parametrically efficient estimator for a sharp bound on the value function under unobserved confounding. Our estimator has three advantages: (1) Unlike existing works, our estimator avoids unstable minimax optimization based on inverse propensity weighted outcomes. (2) Our estimator is semi-parametrically efficient. (3) We prove that our estimator leads to the optimal confounding-robust policy. Finally, we extend our theory to the related task of policy improvement under unobserved confounding, i.e., when a baseline policy such as the standard of care is available. We show in experiments with synthetic and real-world data that our method outperforms simple plug-in approaches and existing baselines. Our method is highly relevant for decision-making where unobserved confounding can be problematic, such as in healthcare and public policy.

Recent flow matching models for text-to-image generation have achieved remarkable quality, yet their integration with reinforcement learning for human preference alignment remains suboptimal, hindering fine-grained reward-based optimization. We observe that the key impediment to effective GRPO training of flow models is the temporal uniformity assumption in existing approaches: sparse terminal rewards with uniform credit assignment fail to capture the varying criticality of decisions across generation timesteps, resulting in inefficient exploration and suboptimal convergence. To remedy this shortcoming, we introduce TempFlow-GRPO (Temporal Flow-GRPO), a principled GRPO framework that captures and exploits the temporal structure inherent in flow-based generation. TempFlow-GRPO introduces three key innovations: (i) a trajectory branching mechanism that provides process rewards by concentrating stochasticity at designated branching points, enabling precise credit assignment without requiring specialized intermediate reward models; (ii) a noise-aware weighting scheme that modulates policy optimization according to the intrinsic exploration potential of each timestep, prioritizing learning during high-impact early stages while ensuring stable refinement in later phases; and (iii) a seed group strategy that controls for initialization effects to isolate exploration contributions. These innovations endow the model with temporally-aware optimization that respects the underlying generative dynamics, leading to state-of-the-art performance in human preference alignment and text-to-image benchmarks.

We present SSD-GS, a physically-based relighting framework built upon 3D Gaussian Splatting (3DGS) that achieves high-quality reconstruction and photorealistic relighting under novel lighting conditions. In physically-based relighting, accurately modeling light-material interactions is essential for faithful appearance reproduction. However, existing 3DGS-based relighting methods adopt coarse shading decompositions, either modeling only diffuse and specular reflections or relying on neural networks to approximate shadows and scattering. This leads to limited fidelity and poor physical interpretability, particularly for anisotropic metals and translucent materials. To address these limitations, SSD-GS decomposes reflectance into four components: diffuse, specular, shadow, and subsurface scattering. We introduce a learnable dipole-based scattering module for subsurface transport, an occlusion-aware shadow formulation that integrates visibility estimates with a refinement network, and an enhanced specular component with an anisotropic Fresnel-based model. Through progressive integration of all components during training, SSD-GS effectively disentangles lighting and material properties, even for unseen illumination conditions, as demonstrated on the challenging OLAT dataset. Experiments demonstrate superior quantitative and perceptual relighting quality compared to prior methods and pave the way for downstream tasks including controllable light source editing and interactive scene relighting.

Reward design is of great importance for solving complex tasks with reinforcement learning. Recent studies have explored using image-text similarity produced by vision-language models (VLMs) to augment rewards of a task with visual feedback. A common practice linearly adds VLM scores to task or success rewards without explicit shaping, potentially altering the optimal policy. Moreover, such approaches, often relying on single static images, struggle with tasks whose desired behavior involves complex, dynamic motions spanning multiple visually different states. Furthermore, single viewpoints can occlude critical aspects of an agent's behavior. To address these issues, this paper presents Multi-View Video Reward Shaping (MVR), a framework that models the relevance of states regarding the target task using videos captured from multiple viewpoints. MVR leverages video-text similarity from a frozen pre-trained VLM to learn a state relevance function that mitigates the bias towards specific static poses inherent in image-based methods. Additionally, we introduce a state-dependent reward shaping formulation that integrates task-specific rewards and VLM-based guidance, automatically reducing the influence of VLM guidance once the desired motion pattern is achieved. We confirm the efficacy of the proposed framework with extensive experiments on challenging humanoid locomotion tasks from HumanoidBench and manipulation tasks from MetaWorld, verifying the design choices through ablation studies.

Generating realistic human-object interaction (HOI) animations remains challenging because it requires jointly modeling dynamic human actions and diverse object geometries. Prior diffusion-based approaches often rely on handcrafted contact priors or human-imposed kinematic constraints to improve contact quality. We propose a data-driven alternative in which guidance emerges from the denoising pace itself, reducing dependence on manually designed priors. Building on diffusion forcing, we factor the representation into modality-specific components and assign individualized noise levels with asynchronous denoising schedules. In this paradigm, cleaner components guide noisier ones through cross-attention, yielding guidance without auxiliary classifiers. We find that this data-driven guidance is inherently contact-aware, and can be further enhanced when training is augmented with a broad spectrum of synthetic object geometries, encouraging invariance of contact semantics to geometric diversity. Extensive experiments show that pace-induced guidance more effectively mirrors the benefits of contact priors than conventional classifier-free guidance, while achieving higher contact fidelity, more realistic HOI generation, and stronger generalization to unseen objects and tasks.

This work presents DCFlow, a novel self-supervised cross-modal flow estimation framework that integrates a decoupled optimization strategy and a cross-modal consistency constraint. Unlike previous unsupervised approaches that implicitly learn flow estimation solely from appearance similarity, we introduce a decoupled optimization strategy with task-specific supervision to address modality discrepancy and geometric misalignment distinctly. This is achieved by collaboratively training a modality transfer network and a flow estimation network. To enable reliable motion supervision without ground-truth flow, we propose a geometry-aware data synthesis pipeline combined with an outlier-robust loss. Additionally, we introduce a cross-modal consistency constraint to jointly optimize both networks, significantly improving flow prediction accuracy. For evaluation, we construct a comprehensive cross-modal flow benchmark by repurposing public datasets. Experimental results demonstrate that DCFlow can be integrated with various flow estimation networks and achieves state-of-the-art performance among unsupervised approaches.

Test-Time Adaptation (TTA) adapts a deployed model during online inference to mitigate the impact of domain shift. While achieving strong accuracy, most existing methods rely on backpropagation, which is memory and computation intensive, making them unsuitable for resource-constrained devices. Recent attempts to reduce this overhead often suffer from high latency or are tied to specific architectures such as ViT-only or CNN-only. In this work, we revisit domain shift from an embedding perspective. Our analysis reveals that domain shift induces three distinct structural changes in the embedding space: translation (mean shift), scaling (variance shift), and rotation (covariance shift). Based on this insight, we propose Progressive Embedding Alignment (PEA), a backpropagation-free and architecture-agnostic TTA approach. By applying a novel covariance alignment procedure at each intermediate layer, PEA efficiently corrects the embedding distortions with only two forward passes. Extensive experiments demonstrate that PEA achieves state-of-the-art performance in both accuracy and efficiency, while also proving versatile across different architectures including ViTs and CNNs.

Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for enhancing Large Language Models (LLMs) on complex reasoning tasks. Yet current methods face an exploration dilemma: standard RL struggles to escape the local optima of pre-trained LLMs’ sharply peaked initial policies, boosting single-solution accuracy (pass@1) but suppressing solution diversity and multi-solution performance (pass@k). As a result, RLVR often distills existing capabilities rather than discovering new reasoning strategies. We address this with a Risk-Sensitive Reinforcement Learning framework. By adopting a risk-seeking objective that interpolates between mean and maximum rewards, we derive a novel Risk-Sensitive GRPO (RS-GRPO) algorithm that emphasizes hard prompts to drive exploration. Across six mathematical reasoning benchmarks and five LLMs, RS-GRPO consistently improves pass@k performance while enhancing or maintaing pass@1.

Whereas most brain–computer interface research has focused on decoding neural signals into behavior or intent, the reverse challenge—using controlled stimuli to steer brain activity—remains far less understood, particularly in the visual domain. However, designing images that consistently elicit desired neural responses is difficult: subjective states lack clear quantitative measures, and EEG feedback is both noisy and non-differentiable. We introduce MindPilot, the first closed-loop framework that uses EEG signals as optimization feedback to guide naturalistic image generation. Unlike prior work limited to invasive settings or low-level flicker stimuli, MindPilot leverages non-invasive EEG with natural images, treating the brain as a black-box function and employing a pseudo-model guidance mechanism to iteratively refine images without requiring explicit rewards or gradients. We validate MindPilot in both simulation and human experiments, demonstrating (i) efficient retrieval of semantic targets, (ii) closed-loop optimization of EEG features, and (iii) human-subject validations in mental matching and emotion regulation tasks. Our results establish the feasibility of EEG-guided image synthesis and open new avenues for non-invasive closed-loop brain modulation, bidirectional brain–computer interfaces, and neural signal–guided generative modeling.

The bird’s-eye view (BEV) representation enables multi-sensor features to be fused within a unified space, serving as the primary approach for achieving comprehensive multi-task perception. However, the discrete grid representation of BEV leads to significant detail loss and limits feature alignment and cross-modal information interaction in multimodal fusion perception. In this work, we break from the conventional BEV paradigm and propose a new universal framework for multi-task multi-modal fusion based on 3D Gaussian representation. This approach naturally unifies multi-modal features within a shared and continuous 3D Gaussian space, effectively preserving edge and fine texture details. To achieve this, we design a novel forward-projection-based multi-modal Gaussian initialization module and a shared cross-modal Gaussian encoder that iteratively updates Gaussian properties based on an attention mechanism. GaussianFusion is inherently a task-agnostic model, with its unified Gaussian representation naturally supporting various 3D perception tasks. Extensive experiments demonstrate the generality and robustness of GaussianFusion. On the nuScenes dataset, it outperforms the 3D object detection baseline BEVFusion by 2.6 NDS. Its variant surpasses GaussFormer on 3D semantic occupancy with 1.55 mIoU improvement while using only 30% of the Gaussians and achieving a 450% speedup.

Adapting large-scale flow and diffusion models to downstream tasks through reward optimization is essential for their adoption in real-world applications, including scientific discovery and image generation. While recent fine-tuning methods based on reinforcement learning and stochastic optimal control achieve compelling performance, they face severe scalability challenges due to high memory demands that scale with model complexity. In contrast, methods that disentangle reward adaptation from base model complexity, such as Classifier Guidance (CG), offer flexible control over computational resource requirements. However, CG suffers from limited reward expressivity and a train-test distribution mismatch due to its offline nature. To overcome the limitations of fine-tuning methods and CG, we propose Value Matching (VM), an online algorithm for learning the value function within an optimal control setting. VM provides tunable memory and compute demands through flexible value network complexity, supports optimization of non-differentiable rewards, and operates on-policy, which enables going beyond the data distribution to discover high-reward regions. Experimentally, we evaluate VM across image generation and molecular design tasks. We demonstrate improved stability and sample efficiency over CG and achieve comparable performance to fine-tuning approaches while requiring less than 5% of their memory usage.

We study the problem of training a risk-sensitive reinforcement learning (RL) agent through imitation learning (IL). Unlike standard IL, our goal is not only to train an agent that matches the expert’s expected return (i.e., its *average performance*) but also its *risk attitude* (i.e., other features of the return distribution, such as variance). We propose a general formulation of the risk-sensitive IL problem in which the objective is to match the expert’s return distribution in Wasserstein distance. We focus on the tabular setting and assume the expert’s reward is *known*. After demonstrating the limited expressivity of Markovian policies for this task, we introduce an efficient and sufficiently expressive subclass of non-Markovian policies tailored to it. Building on this subclass, we develop two provably efficient algorithms—RS-BC and RS-KT —for solving the problem when the transition model is unknown and known, respectively. We show that RS-KT achieves substantially lower sample complexity than RS-BC by exploiting dynamics information. We further demonstrate the sample efficiency of return distribution matching in the setting where the expert’s reward is *unknown* by designing an oracle-based variant of RS-KT. Finally, we complement our theoretical analysis of RS-KT and RS-BC with numerical simulations, highlighting both their sample efficiency and the advantages of non-Markovian policies over standard sample-efficient IL algorithms.

Text embeddings are essential components in modern NLP pipelines. Although numerous embedding models have been proposed, no single model consistently dominates across domains and tasks. This variability motivates the use of ensemble techniques to combine complementary strengths. However, most existing ensemble methods operate on deterministic embeddings and fail to account for model-specific uncertainty, limiting their robustness and reliability in downstream applications. To address these limitations, we propose Uncertainty-driven Embedding Convolution (UEC). UEC first transforms deterministic embeddings into probabilistic ones in a post-hoc manner. It then computes adaptive ensemble coefficients based on embedding uncertainty, derived from a principled surrogate-loss formulation. Additionally, UEC employs an uncertainty-aware similarity function that directly incorporates uncertainty into the similarity scoring, providing a theoretically grounded and efficient surrogate to distributional distances. Extensive experiments on diverse benchmarks demonstrate that UEC consistently improves both performance and robustness by leveraging principled uncertainty modeling.

Reconstructing human motion and its surrounding environment is crucial for understanding human-scene interaction and predicting human movements in the scene. While much progress has been made in capturing human-scene interaction in constrained environments, those prior methods can hardly reconstruct the natural and diverse human motion and scene context from web videos. In this work, we propose JOSH, a novel optimization-based method for 4D human-scene reconstruction in the wild from monocular videos. Compared to prior works that perform separate optimization of the human, the camera, and the scene, JOSH leverages the human-scene contact constraints to jointly optimize all parameters in a single stage. Experiment results demonstrate that JOSH significantly improves 4D human-scene reconstruction, global human motion estimation, and dense scene reconstruction by utilizing the joint optimization of scene geometry, human motion, and camera poses. Further studies show that JOSH can enable scalable training of end-to-end global human motion models on extensive web data, highlighting its robustness and generalizability. The code and model will be publicly available.

Normative and task-driven theories offer powerful top-down explanations for biological systems, yet the goals of quantitatively arbitrating between competing theories, and utilizing them as inductive biases to improve data-driven fits of real biological datasets are prohibitively laborious, and often impossible. To this end, we introduce a Bayesian meta-learning framework designed to automatically convert raw functional predictions from normative theories into tractable probabilistic models. We employ adaptive deep kernel Gaussian processes, meta-learning a kernel on synthetic data generated from a normative theory. This Theory-Informed Kernel specifies a probabilistic model representing the predictions of a given theoretical model: usable for both fitting data and rigorously validating the theory. As a demonstration, we apply our framework to the early visual system, using efficient coding as our normative theory. We show improved response prediction accuracy in ex vivo recordings of mouse retinal ganglion cells stimulated by natural scenes compared to conventional data-driven baselines, while providing accurate uncertainty estimates and interpretable representations. Using exact Bayesian model selection, we also show that our informed kernel can accurately infer the degree of theory-match from data, confirming faithful encapsulation of theory structure. This work provides a more general, scalable, and automated approach for integrating theoretical knowledge into data-driven scientific inquiry in neuroscience and beyond.