ICLR 2026 Papers

Recent advancements in multimodal slow-thinking systems have demonstrated remarkable performance across various visual reasoning tasks. However, their capabilities in text-rich image reasoning tasks remain understudied due to the absence of a dedicated and systematic benchmark. To address this gap, we propose OCR-Reasoning, a novel benchmark designed to systematically assess Multimodal Large Language Models on text-rich image reasoning tasks. Specifically, OCR-Reasoning comprises 1,069 human-annotated examples spanning 6 core reasoning abilities and 18 practical reasoning tasks in text-rich visual scenarios. Unlike existing text-rich image understanding benchmarks that only provide a final answer, this benchmark additionally provides a detailed step-by-step reasoning process. This dual annotation enables the evaluation of both the models' final answers and their reasoning processes, thereby offering a holistic assessment of text-rich reasoning capabilities. By leveraging this benchmark, we conducted a comprehensive evaluation of the latest MLLMs. Our results demonstrate that even the most advanced MLLMs exhibit substantial difficulties in text-rich image reasoning tasks, with none achieving an accuracy above 50\% on our benchmark, indicating that the challenges of text-rich image reasoning are an urgent issue to be addressed. The benchmark and evaluation scripts are available at https://github.com/SCUT-DLVCLab/OCR-Reasoning.

We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propagation from actions to policy parameters when optimizing against some value function. Our key insight is that we can address stable value maximization by avoiding direct optimization over value with the expressive policy and instead construct an on-the-fly RL policy to maximize Q-value. We propose Expressive Policy Optimization (EXPO), a sample-efficient online RL algorithm that utilizes an on-the-fly policy to maximize value with two parameterized policies -- a larger expressive base policy trained with a stable imitation learning objective and a light-weight Gaussian edit policy that edits the actions sampled from the base policy toward a higher value distribution. The on-the-fly policy optimizes the actions from the base policy with the learned edit policy and chooses the value maximizing action from the base and edited actions for both sampling and temporal-difference (TD) backup. Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods both in the setting of fine-tuning a pretrained policy given offline data and in leveraging offline data to train online.

The slow, sequential nature of autoregressive (AR) language models has driven the adoption of parallel decoding methods. However, these non-autoregressive models often sacrifice generation quality because they struggle to model the complex joint distribution of token sequences. To bridge this gap, we introduce Gumbel Distillation, a novel distillation technique that enables parallel decoders to learn this distribution effectively. Our method leverages the Gumbel-Max trick to create a deterministic mapping from a latent Gumbel noise space to the output tokens of a high-performing AR teacher. As a model-agnostic technique, Gumbel Distillation seamlessly integrates with diverse parallel decoding architectures, including MDLM and BD3-LM. Experiments on LM1B and OpenWebText show that Gumbel Distillation substantially improves the generation quality of parallel language models, achieving a 30.0% improvement in MAUVE Score and 10.5% in generative perplexity over MDLM trained on OpenWebText dataset.

Content-aware streaming requires dynamic, chunk-level importance weights to optimize subjective quality of experience (QoE). However, direct human annotation is prohibitively expensive while vision-saliency models generalize poorly. We introduce HiVid, the first framework to leverage Large Language Models (LLMs) as a scalable human proxy to generate high-fidelity weights for both Video-on-Demand (VOD) and live streaming. We address 3 non-trivial challenges: (1) To extend LLMs' limited modality and circumvent token limits, we propose a perception module to assess frames in a local context window, autoregressively building a coherent understanding of the video. (2) For VOD with rating inconsistency across local windows, we propose a ranking module to perform global re-ranking with a novel LLM-guided merge-sort algorithm. (3) For live streaming which requires low-latency, online inference without future knowledge, we propose a prediction module to predict future weights with a multi-modal time series model, which comprises a content-aware attention and adaptive horizon to accommodate asynchronous LLM inference. Extensive experiments show HiVid improves weight prediction accuracy by up to 11.5\% for VOD and 26\% for live streaming over SOTA baselines. Real-world user study validates HiVid boosts streaming QoE correlation by 14.7\%.

Learning distributions over permutations is a fundamental problem in machine learning, with applications in ranking, combinatorial optimization, structured prediction, and data association. Existing methods rely on mixtures of parametric families or neural networks with expensive variational inference procedures. In this work, we propose a novel approach that leverages alternative representations for permutations, including Lehmer codes, Fisher-Yates draws, and Insertion-Vectors. These representations form a bijection with the symmetric group, allowing for unconstrained learning using conventional deep learning techniques, and can represent any probability distribution over permutations. Our approach enables a trade-off between expressivity of the model family and computational requirements. In the least expressive and most computationally efficient case our method subsumes previous families of well established probabilistic models over permutations, including Mallow's and the Repeated Insertion Model. Experiments indicate our method significantly outperforms current approaches on the jigsaw puzzle benchmark, a common task for permutation learning. However, we argue this benchmark is limited in its ability to assess learning probability distributions, as the target is a delta distribution (i.e., a single correct solution exists). We therefore propose two additional benchmarks: learning cyclic permutations and re-ranking movies based on user preference. We show that our method learns non-trivial distributions even in the least expressive mode, while traditional models fail to even generate valid permutations in this setting.

Deep generative models hold great promise for representing complex physical systems, but their deployment is currently limited by the lack of guarantees on the physical plausibility of the generated outputs. Ensuring that known physical constraints are enforced is therefore critical when applying generative models to scientific and engineering problems. We address this limitation by developing a principled framework for sampling from a target distribution while rigorously satisfying physical constraints. Leveraging the variational formulation of Langevin dynamics, we propose Split Augmented Langevin (SAL), a novel primal-dual sampling algorithm that enforces constraints progressively through variable splitting, with convergence guarantees. While the method is developed theoretically for Langevin dynamics, we demonstrate its effective applicability to diffusion models. In particular, we use constrained diffusion models to generate physical fields satisfying energy and mass conservation laws. We apply our method to diffusion-based data assimilation on a complex physical system, where enforcing physical constraints substantially improves both forecast accuracy and the preservation of critical conserved quantities. We also demonstrate the potential of SAL for challenging feasibility problems in optimal control.

Automatically generating interactive 3D environments is crucial for scaling up robotic data collection in simulation. While prior work has primarily focused on 3D asset placement, it often overlooks the physical relationships between objects (e.g., contact, support, balance, and containment), which are essential for creating complex and realistic manipulation scenarios such as tabletop arrangements, shelf organization, or box packing. Compared to classical 3D layout generation, producing complex physical scenes introduces additional challenges: (a) higher object density and complexity (e.g., a small shelf may hold dozens of books), (b) richer supporting relationships and compact spatial layouts, and (c) the need to accurately model both spatial placement and physical properties. To address these challenges, we propose PhyScensis, an LLM agent-based framework powered by a physics engine, to produce physically plausible scene configurations with high complexity. Specifically, our framework consists of three main components: an LLM agent iteratively proposes assets with spatial and physical predicates; a solver, equipped with a physics engine, realizes these predicates into a 3D scene; and feedback from the solver informs the agent to refine and enrich the configuration. Moreover, our framework preserves strong controllability over fine-grained textual descriptions and numerical parameters (e.g., relative positions, scene stability), enabled through probabilistic programming for stability and a complementary heuristic that jointly regulates stability and spatial relations. Experimental results show that our method outperforms prior approaches in scene complexity, visual quality, and physical accuracy, offering a unified pipeline for generating complex physical scene layouts for robotic manipulation.

Diffusion alignment aims to optimize diffusion models for the downstream objective. While existing methods based on reinforcement learning or direct backpropagation achieve considerable success in maximizing rewards, they often suffer from reward over-optimization and mode collapse. We introduce Diffusion Alignment as Variational Expectation-Maximization (DAV), a framework that formulates diffusion alignment as an iterative process alternating between two complementary phases: the E-step and the M-step. In the E-step, we employ test-time search to generate diverse and reward-aligned samples. In the M-step, we refine the diffusion model using samples discovered by the E-step. We demonstrate that DAV can optimize reward while preserving diversity for both continuous and discrete tasks: text-to-image synthesis and DNA sequence design.

We partially reverse-engineer a convolutional recurrent neural network (RNN) trained with model-free reinforcement learning to play the box-pushing game Sokoban. We find that the RNN stores future moves (plans) as activations in particular channels of the hidden state, which we call *path channels*. A high activation in a particular location means that, when a box is in that location, it will get pushed in the channel's assigned direction. We examine the convolutional kernels between path channels and find that they encode the change in position resulting from each possible action, thus representing part of a learned *transition model*. The RNN constructs plans by starting at the boxes and goals. These kernels, *extend* activations in path channels forwards from boxes and backwards from the goal. Negative values are placed in channels at obstacles. This causes the extension kernels to propagate the negative value in reverse, thus pruning the last few steps and letting an alternative plan emerge; a form of backtracking. Our work shows that, a precise understanding of the plan representation allows us to directly understand the bidirectional planning-like algorithm learned by model-free training in more familiar terms.

Image restoration in real-world conditions is highly challenging due to heterogeneous degradations such as haze, noise, shadows, and blur. Existing diffusion-based methods remain limited: conditional generation struggles to balance fidelity and realism, inversion-based approaches accumulate errors, and posterior sampling requires a known forward operator that is rarely available. We introduce **LearnIR**, a learnable diffusion posterior sampling framework that eliminates this dependency by training a lightweight model to directly predict gradient correction distributions, enabling *Diffusion Posterior Sampling Correction (DPSC)* that maintains consistency with the true image distribution during sampling. In addition, a *Dynamic Resolution Module (DRM)* dynamically adjusts resolution to preserve global structures in early stages and refine fine textures later, while avoiding the need for a pretrained VAE. Experiments on ISTD, O-HAZE, HazyDet, REVIDE, and our newly constructed FaceShadow dataset show that LearnIR achieves state-of-the-art performance in PSNR, SSIM, and LPIPS.

Large language models (LLMs) can exhibit *concept-conditioned semantic divergence*: common high-level cues (e.g., ideologies, public figures) elicit unusually uniform, stance-like responses that evade token-trigger audits. This behavior falls in a blind spot of current safety evaluations, yet carries major societal stakes, as such concept cues can steer content exposure at scale. We formalize this phenomenon and present **RAVEN** (**R**esponse **A**nomaly **V**igilance), a black-box audit that flags cases where a model is simultaneously highly certain and atypical among peers by coupling *semantic entropy* over paraphrastic samples with *cross-model disagreement*. In a controlled LoRA fine-tuning study, we implant a concept-conditioned stance using a small biased corpus, demonstrating feasibility without rare token triggers. Auditing five LLM families across twelve sensitive topics (360 prompts per model) and clustering via bidirectional entailment, RAVEN surfaces recurrent, model-specific divergences in 9/12 topics. Concept-level audits complement token-level defenses and provide a practical early-warning signal for release evaluation and post-deployment monitoring against propaganda-like influence.

We propose **FROST**, an attention-aware method for efficient reasoning. Unlike traditional approaches, FROST leverages attention weights to prune uncritical reasoning paths, yielding shorter and more reliable reasoning trajectories. Methodologically, we introduce the concept of *reasoning outliers* and design an attention-based mechanism to remove them. Theoretically, FROST preserves and enhances the model’s reasoning capacity while eliminating outliers at the sentence level. Empirically, we validate FROST on four benchmarks using two strong reasoning models (Phi-4-Reasoning and GPT-oss-20B), outperforming state-of-the-art methods such as TALE and ThinkLess. Notably, FROST achieves an average **69.68%** reduction in token usage and a **26.70%** improvement in accuracy over the base model. Furthermore, in evaluations of attention outlier metrics, FROST reduces the maximum infinity norm $\||\mathbf{x}\||_{\infty}$ by **15.97%** and the average kurtosis by **91.09%** compared to the base model.

Imitation learning for end-to-end driving trains policies only on expert demonstrations. Once deployed in a closed loop, such policies lack recovery data: small mistakes cannot be corrected and quickly compound into failures. A promising direction is to generate alternative viewpoints and trajectories beyond the logged path. Prior work explores photorealistic digital twins via neural rendering or game engines, but these methods are prohibitively slow and costly, and thus mainly used for evaluation. In this work, we argue that photorealism is unnecessary for training end-to-end planners. What matters is semantic fidelity and scalability: driving depends on geometry and dynamics, not textures or lighting. Motivated by this, we propose 3D Rasterization, which replaces costly rendering with lightweight rasterization of annotated primitives, enabling augmentations such as counterfactual recovery maneuvers and cross-agent view synthesis. To transfer these synthetic views effectively to real-world deployment, we introduce a Raster-to-Real (R2R) feature-space alignment that bridges the sim-to-real gap at the representation level. Together, these components form the Rasterization Augmented Planning (RAP) pipeline, a scalable data augmentation framework for planning. RAP achieves state-of-the-art closed-loop robustness and long-tail generalization, ranking 1st on four major benchmarks: NAVSIM v1/v2, Waymo Open Dataset Vision-based E2E Driving, and Bench2Drive. Our results demonstrate that lightweight rasterization with feature alignment suffices to scale end-to-end training, offering a practical alternative to photorealistic rendering. Project page: https://alan-lanfeng.github.io/RAP/.

Large language models (LLMs) can be prompted with specific styles (e.g., formatting responses as lists), including in malicious queries. Prior jailbreak research mainly augments these queries with additional string transformations to maximize attack success rate (ASR). However, the impact of style patterns in the original queries that are semantically irrelevant to the malicious intent remains unclear. In this work, we seek to understand whether style patterns compromise LLM safety, how superficial style alignment increases model vulnerability, and how best to mitigate these risks during alignment. We first define ASR inflation as the increase in ASR due to style patterns in existing jailbreak benchmark queries. By evaluating $32$ LLMs across seven benchmarks, we find that nearly all models exhibit ASR inflation. Notably, the inflation correlates with an LLM's relative attention to style patterns, which also overlap more with its instruction-tuning data when inflation occurs. We then investigate superficial style alignment, and find that fine-tuning with specific styles makes LLMs more vulnerable to jailbreaks of those same styles. Finally, we propose SafeStyle, a defense strategy that incorporates a small amount of safety training data augmented to match the distribution of style patterns in the fine-tuning data. Across three LLMs, six fine-tuning style settings, and two real-world instruction-tuning datasets, SafeStyle consistently outperforms baselines in maintaining LLM safety.

AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 1.5: a carefully curated hard benchmark composed of 74 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 50% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work.

LLM inference often generates a batch of candidates for a prompt and selects one via strategies like majority voting or Best-of- N (BoN). For difficult tasks, this single-shot selection often underperforms. Consequently, evaluations commonly report Pass@$k$: the agent may submit up to $k$ responses, and only the best of them is used when computing regret. Motivated by this, we study inference scaling in the more general Pass@$k$ inference setting, and prove that neither majority voting nor BoN exhibits the desirable scaling with $k$ and the sampling budget $N$. Combining the advantages of majority voting and BoN, we propose a new inference strategy called Best-of-Majority (BoM), with a pivotal step that restricts the candidates to the responses with high frequency in the $N$ samples before selecting the top-$k$ rewards. We prove that when the sampling budget is $N=\tilde\Omega(C^\*)$, the regret of BoM is $O(\epsilon_{\mathrm{opt}}+\sqrt{\epsilon_{\mathrm{RM}}^2C^\*/k})$, where $C^*$ is the coverage coefficient, $\epsilon_{\mathrm{RM}}$ is the estimation error of the reward model, and $\epsilon_{\mathrm{opt}}$ is the estimation error of reward at the optimal response. We further establish a matching lower bound, certifying that our algorithm is minimax optimal. Beyond optimality, BoM has a key advantage: unlike majority voting and BoN, its performance does not degrade when increasing $N$. Experimental results of inference on math problems show BoM outperforming both majority voting and BoN.

We introduce a resource-efficient neural network architecture with zero divergence by design, adapted for high-dimensional problems. Our method is directly applicable to image denoising, for which divergence-free estimators are particularly well-suited for self-supervised learning, in accordance with Stein's unbiased risk estimation theory. Comparisons of our parameterization on popular denoising datasets demonstrate that it retains sufficient expressivity to remain competitive with other divergence-based approaches, while outperforming its counterparts when the noise level is unknown and varies across the training data.

Logical specifications have been shown to help reinforcement learning algorithms in achieving complex tasks. However, when a task is under-specified, agents might fail to learn useful policies. In this work, we explore the possibility of improving coarse-grained logical specifications via an exploration-guided strategy. We propose **AutoSpec**, a framework that searches for a logical specification refinement whose satisfaction implies satisfaction of the original specification, but which provides additional guidance therefore making it easier for reinforcement learning algorithms to learn useful policies. **AutoSpec** is applicable to reinforcement learning tasks specified via the SpectRL specification logic. We exploit the compositional nature of specifications written in SpectRL, and design four refinement procedures that modify the abstract graph of the specification by either refining its existing edge specifications or by introducing new edge specifications. We prove that all four procedures maintain specification soundness, i.e. any trajectory satisfying the refined specification also satisfies the original. We then show how **AutoSpec** can be integrated with existing reinforcement learning algorithms for learning policies from logical specifications. Our experiments demonstrate that **AutoSpec** yields promising improvements in terms of the complexity of control tasks that can be solved, when refined logical specifications produced by **AutoSpec** are utilized.

In this paper, we investigate federated learning for quantile inference under local differential privacy (LDP). We propose an estimator based on local stochastic gradient descent (SGD), whose local gradients are perturbed via a randomized mechanism with global parameters, making the procedure tolerant of communication and storage constraints without compromising statistical efficiency. Although the quantile loss and its corresponding gradient do not satisfy standard smoothness conditions typically assumed in existing literature, we establish asymptotic normality for our estimator as well as a functional central limit theorem. The proposed method accommodates data heterogeneity and allows each server to operate with an individual privacy budget. Furthermore, we construct confidence intervals for the target value through a self‐normalization approach, thereby circumventing the need to estimate additional nuisance parameters. Extensive numerical experiments and real data application validate the theoretical guarantees of the proposed methodology.