ICLR 2026 论文列表

Masked diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive LLMs, offering competitive performance while supporting unique generation capabilities such as inpainting. We explore how inpainting can inform RL algorithm design for dLLMs. Aligning LLMs with reinforcement learning faces an exploration challenge: sparse reward signals and sample waste when models fail to discover correct solutions. While this inefficiency affects LLMs broadly, dLLMs offer a distinctive opportunity—their inpainting ability can guide exploration. We introduce IGPO (Inpainting Guided Policy Optimization), an RL framework that strategically inserts partial ground-truth reasoning traces during online sampling. Unlike providing full solutions, inpainting steers exploration toward promising trajectory spaces while preserving self-generated reasoning, bridging supervised fine-tuning and reinforcement learning. We apply IGPO to group-based optimization methods such as GRPO, where exploration failures cause zero advantages and gradients. IGPO restores meaningful gradients while improving sample efficiency. We also propose supervised fine-tuning on synthetically rewritten concise traces that better align with dLLM generation patterns. With additional techniques including entropy-based filtering, our training recipe yields substantial gains across four mathematical benchmarks—GSM8K, Math500, AMC and Minerva—achieving new state-of-the-art results for full-attention masked dLLMs.

Modeling 3D dynamics is a fundamental problem in multi-body systems across scientific and engineering domains and has important practical implications in trajectory prediction and simulation. While recent GNN-based approaches have achieved strong performance by enforcing geometric symmetries, encoding high-order features or incorporating neural-ODE mechanics, they typically depend on explicitly observed structures and inherently fail to capture the unobserved interactions that are crucial to complex physical behaviors and dynamics mechanism. In this paper, we propose PAINET, a principled SE(3)-equivariant neural architecture for learning all-pair interactions in multi-body systems. The model comprises: (1) a novel physics-inspired attention network derived from the minimization trajectory of an energy function, and (2) a parallel decoder that preserves equivariance while enabling efficient inference. Empirical results on diverse real-world benchmarks, including human motion capture, molecular dynamics, and large-scale protein simulations, show that PAINET consistently outperforms recently proposed models, yielding 4.7% to 41.5% error reductions in 3D dynamics prediction with comparable computation costs in terms of time and memory. Our codes, baseline models and datasets are available at https://anonymous.4open.science/r/PAINET-4668.

We tackle the task of long-form music generation, particularly the challenging \textbf{lyrics-to-song} problem, by introducing \textbf{YuE (乐)}, a family of open-source music generation foundation models. Specifically, YuE scales to trillions of tokens and generates up to five minutes of music while maintaining lyrical alignment, coherent musical structure, and engaging vocal melodies with appropriate accompaniment. It achieves this through \textbf{track-decoupled next-token prediction} to overcome dense mixture signals, and \textbf{structural progressive conditioning} for long-context lyrical alignment. In addition, we redesign the \textbf{in-context learning} technique for music generation, enabling bidirectional content creation, style cloning, and improving musicality. Through extensive evaluation, we demonstrate that YuE matches or even surpasses some of the proprietary systems in musicality and vocal agility (as of 2025-01). We strongly encourage readers to \textbf{listen to our demo}\footnote{\url{https://yue-anonymous.github.io}}.

In arbitrary-order language models, it is an open question how to sample tokens in parallel from the correct joint distribution. With discrete diffusion models, the more tokens they generate in parallel, the less their predicted distributions adhere to the originally learned data distribution, as they rely on a conditional independence assumption that only works with infinitesimally small timesteps. We find that a different class of models, any-subset autoregressive models (AS-ARMs), holds the solution. As implied by the name, AS-ARMs can generate tokens in any order, and in parallel. Moreover, AS-ARMs support parallelized joint probability density estimation, allowing them to correct their own parallel-generated token distributions, via our Any-Subset Speculative Decoding (ASSD) algorithm. ASSD provably enables generation of tokens from the correct joint distribution, with the number of neural network calls upper bounded by the number of tokens predicted – notably, previous speculative decoding algorithms lack our efficiency guarantee. We empirically verify that ASSD speeds up language generation, without sacrificing quality. Furthermore, we provide a mathematically justified scheme for training AS-ARMs for generation, and show that AS-ARMs achieve state-of-the-art performance among sub-200M parameter models on infilling benchmark tasks, and nearly match the performance of models 50X larger on code generation. Our theoretical and empirical results indicate that the once-forgotten AS-ARMs are a promising direction of language modeling.

Recent works have explored reference-based super-resolution (RefSR) to mitigate hallucinations in diffusion-based image restoration. A key challenge is that real-world degradations make correspondences between low-quality (LQ) inputs and reference (Ref) images unreliable, requiring adaptive control of reference usage. Existing methods either ignore LQ–Ref correlations or rely on brittle explicit matching, leading to over-reliance on misleading references or under-utilization of valuable cues. To address this, we propose Ada-RefSR, a single-step diffusion framework guided by a "Trust but Verify " principle: reference information is leveraged when reliable and suppressed otherwise. Its core component, Adaptive Implicit Correlation Gating (AICG), employs learnable summary tokens to distill dominant reference patterns and capture implicit correlations with LQ features. Integrated into the attention backbone, AICG provides lightweight, adaptive regulation of reference guidance, serving as a built-in safeguard against erroneous fusion. Experiments on multiple datasets demonstrate that Ada-RefSR achieves a strong balance of fidelity, naturalness, and efficiency, while remaining robust under varying reference alignment. Code and models are available at https://github.com/vivoCameraResearch/AdaRefSR.

We study density estimation from pairwise comparisons, motivated by expert knowledge elicitation and learning from human feedback. We relate the unobserved target density to a tempered winner density (marginal density of preferred choices), learning the winner's score via score-matching. This allows estimating the target by `de-tempering' the estimated winner density's score. We prove that the score vectors of the belief and the winner density are collinear, linked by a position-dependent tempering field. We give analytical formulas for this field and propose an estimator for it under the Bradley-Terry model. Using a diffusion model trained on tempered samples generated via score-scaled annealed Langevin dynamics, we can learn complex multivariate belief densities of simulated experts, from only hundreds to thousands of pairwise comparisons.

Recent studies have extended Reinforcement Learning with Verifiable Rewards (RLVR) to autoregressive (AR) visual generation and achieved promising progress. However, existing methods typically apply uniform optimization across all image tokens, while the varying contributions of different image tokens for RLVR's training remain unexplored. In fact, the key obstacle lies in how to identify more critical image tokens during AR generation and implement effective token-wise optimization for them. To tackle this challenge, we propose $\textbf{G}$roup $\textbf{C}$ritical-token $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{GCPO}$), which facilitates effective policy optimization on critical tokens. We identify the critical tokens in RLVR-based AR generation from three perspectives, specifically: $\textbf{(1)}$ Causal dependency: early tokens fundamentally determine the later tokens and final image effect due to unidirectional dependency; $\textbf{(2)}$ Entropy-induced spatial structure: tokens with high entropy gradients correspond to image structure and bridges distinct visual regions; $\textbf{(3)}$ RLVR-focused token diversity: tokens with low visual similarity across a group of sampled images contribute to richer token-level diversity. For these identified critical tokens, we further introduce a dynamic token-wise advantage weight to encourage exploration, based on confidence divergence between the policy model and reference model. By leveraging 30\% of the image tokens, GCPO achieves better performance than GRPO with full tokens. Extensive experiments on multiple text-to-image benchmarks for both AR models and unified multimodal models demonstrate the effectiveness of GCPO for AR visual generation.

In continual learning, knowledge must be preserved and re-used between tasks, requiring a balance between maintaining good transfer to future tasks and minimizing forgetting of previously learned ones. As several practical algorithms have been devised to address the continual learning setting, the natural question of providing reliable risk certificates has also been raised. Although there are results for specific settings and algorithms on the behavior of memory stability, generally applicable upper bounds on learning plasticity are few and far between. In this work, we extend existing PAC-Bayes bounds for online learning and time-uniform offline learning to the continual learning setting. We derive general upper bounds on the cumulative generalization loss applicable for any task distribution and learning algorithm as well as oracle bounds for Gibbs posteriors and compare their effectiveness for several different task distributions. We demonstrate empirically that our approach yields non-vacuous bounds for several continual learning problems in vision, as well as tight oracle bounds on linear regression tasks. To the best of our knowledge, this is the first general upper bound on learning plasticity for continual learning.

Reasoning LLMs are trained to verbalize their thinking process, yielding strong gains on complex tasks. This transparency also opens a promising direction: multiple reasoners can directly collaborate on each other's thinking on a shared trajectory, yielding better inference efficiency and exploration. A key prerequisite, however, is the ability to assess usefulness and build on another model's partial thinking —we call this *off-trajectory reasoning*. Our paper investigates a critical question: can standard *solo-reasoning* training pipelines yield desired *off-trajectory* behaviors? We propose twin tests that capture the two extremes of the off-trajectory spectrum, namely **Recoverability**, which tests whether LLMs can backtrack from ``distractions'' induced by misleading reasoning traces, and **Guidability**, which tests their ability to build upon correct reasoning from stronger collaborators. Our study evaluates 15 open-weight LLMs (1.5B—32B) and reveals a counterintuitive finding — "stronger" LLMs on benchmarks are often more fragile under distraction. Moreover, all models tested fail to effectively leverage guiding steps from collaborators on problems beyond their inherent capabilities, with solve rates remaining under 9.2%. Finally, we conduct control studies to isolate the effects of three factors in post-training on these behaviors: the choice of distillation teacher, the use of RL, and data selection strategy. Our results provide actionable insights for training natively strong reasoning collaborators; e.g., we find that sub-optimal recoverability behaviors of teacher models are transferred to distilled students even if the distilled data trajectories are correct. Taken together, this work lays the groundwork for evaluating multi-model collaborations under shared reasoning, while revealing limitations of off-the-shelf reasoning LLMs.

Diffusion large language models (dLLMs) have recently drawn considerable attention within the research community as a promising alternative to autoregressive generation, offering parallel token prediction and lower inference latency. Yet, their parallel decoding potential remains largely underexplored, as existing open-source models still require nearly token-length decoding steps to ensure performance. To address this, we introduce dParallel, a simple and effective method that unlocks the inherent parallelism of dLLMs for fast sampling. We identify that the key bottleneck to parallel decoding arises from the sequential certainty convergence for masked tokens. Building on this insight, we introduce the core of our approach: certainty-forcing distillation, a novel training strategy that distills the model to follow its original sampling trajectories while enforcing it to achieve high certainty on masked tokens more rapidly and in parallel. Extensive experiments demonstrate that our method can dramatically reduce the number of decoding steps while maintaining performance. When applied to the LLaDA-8B-Instruct model, dParallel reduces decoding steps from 256 to 30 on GSM8K, achieving an 8.5× speedup without performance degradation. On the MBPP benchmark, it cuts decoding steps from 256 to 24, resulting in a 10.5× speedup while maintaining accuracy.

Universal image restoration is a critical task in low-level vision, requiring the model to remove various degradations from low-quality images to produce clean images with rich detail. The challenges lie in sampling the distribution of high-quality images and adjusting the outputs on the basis of the degradation. This paper presents a novel approach, Bridging Degradation discrimination and Generation (BDG), which aims to address these challenges concurrently. First, we propose the Multi-Angle and multi-Scale Gray Level Co-occurrence Matrix (MAS-GLCM) and demonstrate its effectiveness in performing fine-grained discrimination of degradation types and levels. Subsequently, we divide the diffusion training process into three distinct stages: generation, bridging, and restoration. The objective is to preserve the diffusion model's capability of restoring rich textures while simultaneously integrating the discriminative information from the MAS-GLCM into the restoration process. This enhances its proficiency in addressing multi-task and multi-degraded scenarios. Without changing the architecture, BDG achieves significant performance gains in all-in-one restoration and real-world super-resolution tasks, primarily evidenced by substantial improvements in fidelity without compromising perceptual quality.

Compositional Zero-Shot Learning (CZSL) investigates compositional generalization capacity to recognize unknown state-object pairs based on learned primitive concepts. Existing CZSL methods typically derive primitives features through a simple composition-prototype mapping, which is suboptimal for a set of individuals that can be divided into distinct semantic subsets. Moreover, the one-to-all cross-modal primitives matching neglects compositional divergence within identical states or objects, limiting fine-grained image-composition alignment. In this study, we propose EVA, a Mixture-of-Experts Framework for Semantic Variant Alignment. Specifically, we introduce domain-expert adaption, leveraging multiple experts to achieve token-aware learning and model high-quality primitive representations. To enable accurate compositional generalization, we further present semantic variant alignment to select semantically relevant representation for image-primitives matching. Our method significantly outperforms other state-of-the-art CZSL methods on three popular benchmarks in both closed- and open-world settings, demonstrating the efficacy of the proposed insight.

Multimodal large language models (MLLMs) have significantly advanced the integration of visual and textual understanding. However, their ability to generate code from multimodal inputs remains limited. In this work, we introduce VisCodex, a unified framework that seamlessly merges vision and coding language models to empower MLLMs with strong multimodal code generation abilities. Leveraging a task vector-based model merging technique, we integrate a state-of-the-art coding LLM into a strong vision-language backbone, while preserving both visual comprehension and advanced coding skills. To support training and evaluation, we introduce the Multimodal Coding Dataset (MCD), a large-scale and diverse collection of 598k samples, including high-quality HTML code, chart image-code pairs, image-augmented StackOverflow QA, and algorithmic problems. Furthermore, we propose InfiBench-V, a novel and challenging benchmark specifically designed to assess models on visually-rich, real-world programming questions that demand a nuanced understanding of both textual and visual contexts. Extensive experiments show that VisCodex achieves state-of-the-art performance among open-source MLLMs and approaches proprietary models like GPT-4o, highlighting the effectiveness of our model merging strategy and new datasets.

Active perception is a fundamental skill that enables us humans to deal with uncertainty in our inherently partially observable environment. For senses such as touch, where the information is sparse and local, active perception becomes crucial. In recent years, active perception has emerged as an important research domain in robotics. However, current methods are often bound to specific tasks or make strong assumptions, which limit their generality. To address this gap, this work introduces APPLE (Active Perception Policy Learning) - a novel framework that leverages reinforcement learning (RL) to address a range of different active perception problems. APPLE jointly trains a transformer-based perception module and decision-making policy with a unified optimization objective, learning how to actively gather information. By design, APPLE is not limited to a specific task and can, in principle, be applied to a wide range of active perception problems. We evaluate two variants of APPLE across different tasks, including tactile exploration problems from the Tactile MNIST benchmark. Experiments demonstrate the efficacy of APPLE, achieving high accuracies on both regression and classification tasks. These findings underscore the potential of APPLE as a versatile and general framework for advancing active perception in robotics.

Differential Privacy (DP) is becoming central to large-scale training as privacy regulations tighten. We revisit how DP noise interacts with *adaptivity* in optimization through the lens of *stochastic differential equations*, providing the first SDE-based analysis of private optimizers. Focusing on DP-SGD and DP-SignSGD under per-example clipping, we show a sharp contrast under fixed hyperparameters: DP-SGD converges at a privacy-utility trade-off $O(1/\varepsilon^2)$ with speed independent of $\varepsilon$, while DP-SignSGD converges at a speed *linear* in $\varepsilon$ with a $O(1/\varepsilon)$ trade-off, dominating in high-privacy or high-noise regimes. Under optimal learning rates, both methods reach comparable theoretical asymptotic performance; however, the optimal learning rate of DP-SGD scales linearly with $\varepsilon$, while that of DP-SignSGD is essentially $\varepsilon$-independent. This makes adaptive methods far more practical, as their hyperparameters transfer across privacy levels with little or no re-tuning. Empirical results confirm our theory across training and test metrics, and extend from DP-SignSGD to DP-Adam.

Although diffusion models have achieved strong results in decision-making tasks, their slow inference speed remains a key limitation. While consistency models offer a potential solution, existing applications to decision-making either struggle with suboptimal demonstrations under behavior cloning or rely on complex concurrent training of multiple networks under the actor-critic framework. In this work, we propose a novel approach to consistency distillation for offline reinforcement learning that directly incorporates reward optimization into the distillation process. Our method achieves single-step sampling while generating higher-reward action trajectories through decoupled training and noise-free reward signals. Empirical evaluations on the Gym MuJoCo, FrankaKitchen, and long horizon planning benchmarks demonstrate that our approach can achieve a $9.7$% improvement over previous state-of-the-art while offering up to $142\times$ speedup over diffusion counterparts in inference time.

Diffusion Transformers (DiTs) typically use the same patch size for $\operatorname{Patchify}$ across timesteps, enforcing a constant token budget across timesteps. In this paper, we introduce Pyramidal Patchification Flow (PPFlow), which reduces the number of tokens for high-noise timesteps to improve the sampling efficiency. The idea is simple: use larger patches at higher-noise timesteps and smaller patches at lower-noise timesteps. The implementation is easy: share the DiT's transformer blocks across timesteps, and learn separate linear projections for different patch sizes in $\operatorname{Patchify}$ and $\operatorname{Unpatchify}$. Unlike Pyramidal Flow that operates on pyramid representations,, our approach operates over full latent representations, eliminating trajectory ``jump points'', and thus avoiding re-noising tricks for sampling. Training from pretrained SiT-XL/2 requires only $+8.9\%$ additional training FLOPs and delivers $2.02\times$ denoising speedups with image generation quality kept; training from scratch achieves comparable sampling speedup, e.g., $2.04\times$ speedup in SiT-B. Training from text-to-image model FLUX.1, PPFlow can achieve $1.61 - 1.86 \times$ speedup from 512 to 2048 resolution with comparable quality.

Generating (pseudo-)random variates lies at the core of probabilistic machine learning and prediction algorithms and yet remains a major bottleneck due to its high computational and energy cost. In this paper, we introduce a general and scalable sampling strategy that enables fast and energy-efficient random variate generation from arbitrary distributions. Our approach is based on efficient lookup tables combined with a fast index sampling scheme. Using only a handful of fast and energy-efficient compute operations on simple array structures, we achieve superior speed, energy efficiency, and precision at near-optimal entropy cost compared to state-of-the-art techniques. Microbenchmarking our approach with a C implementation shows up to 40\% savings in time and 60\% in energy compared to state-of-the-art approaches. Compared to commonly employed Python samplers we achieve a 100x time improvement.

Recent AIGC advances have rapidly expanded from text-to-image generation toward high-quality multimodal synthesis across video and audio. Within this context, joint audio-video generation (JAVG) has emerged as a fundamental task that produces synchronized and semantically aligned sound and vision from textual descriptions. However, compared with advanced commercial models such as Veo3, existing open-source methods still suffer from limitations in generation quality, temporal synchrony, and alignment with human preferences. To bridge the gap, this paper presents JavisDiT++, a concise yet powerful framework for efficient and effective JAVG. First, we introduce a modality-specific mixture-of-experts (MS-MoE) design that enables cross-modal interaction efficacy while enhancing single-modal generation quality. Then, we propose a temporal-aligned RoPE (TA-RoPE) strategy to achieve explicit, frame-level synchronization between audio and video tokens. Besides, we develop an audio-video direct preference optimization (AV-DPO) method to align model outputs with human preference across quality, consistency, and synchrony dimensions. Built upon Wan2.1-1.3B-T2V, our model achieves state-of-the-art performance merely with around 1M public training entries, significantly outperforming prior approaches in both qualitative and quantitative evaluations. Comprehensive ablation studies have been conducted to validate the effectiveness of our proposed modules. All the code, model, and dataset are released at https://JavisVerse.github.io/JavisDiT2-page.