ICLR 2026 Papers

One-step generators distilled from Masked Diffusion Models (MDMs) compress multiple sampling steps into a single forward pass, enabling efficient text and image synthesis. However, they suffer two key limitations: they inherit modeling bias from the teacher, and their discrete token outputs block gradient flow, preventing post-distillation refinements such as adversarial training, reward-based fine-tuning, and Test-Time Embedding Optimization (TTEO). In this work, we introduce soft embeddings, a simple relaxation that replaces discrete tokens with the expected embeddings under the generator's output distribution. Soft embeddings preserve representation fidelity for one-step discrete generator while providing a fully differentiable continuous surrogate that is compatible with teacher backbones and tokenizer decoders while cause minimum bias. Integrating soft embeddings into the Di[M]O \citep{zhu2025di} distillation framework (denoted Soft-Di[M]O) makes one-step generators end-to-end trainable and enables straightforward application of GAN-based refinement, differentiable reward fine-tuning, and TTEO. Empirically, across multiple MDM teachers (e.g., MaskBit \citep{weber2024maskbit}, MaskGen \citep{kim2025democratizing}), Soft-Di[M]O achieves state-of-the-art one-step results: improved class-to-image performance, a one-step FID of 1.56 on ImageNet-256 with GAN-based refinement, along with higher than teacher GenEval \citep{ghosh2023geneval} and HPS \citep{wu2023human} scores on text-to-image with reward fine-tuning, and further gains from TTEO.

Progress in long-context reasoning for large language models (LLMs) has lagged behind other recent advances. This gap arises not only from the intrinsic difficulty of processing long texts, but also from the scarcity of reliable human annotations and programmatically verifiable reward signals. In this paper, we propose SPELL, a multi-role self-play reinforcement learning framework that enables scalable, label-free optimization for long-context reasoning. SPELL integrates three cyclical roles—questioner, responder, and verifier—within a single model to enable continual self-improvement. The questioner generates questions from raw documents paired with reference answers; the responder learns to solve these questions based on the documents; and the verifier evaluates semantic equivalence between the responder’s output and the questioner's reference answer, producing reward signals to guide continual training. To stabilize training, we introduce an automated curriculum that gradually increases document length and a reward function that adapts question difficulty to the model’s evolving capabilities. Extensive experiments on six long-context benchmarks show that SPELL consistently improves performance across diverse LLMs and outperforms equally sized models fine-tuned on large-scale annotated data. Notably, SPELL achieves an average 7.6-point gain in pass@8 on the strong reasoning model Qwen3-30B-A3B-Thinking, raising its performance ceiling and showing promise for scaling to even more capable models. Our code is available at https://github.com/Tongyi-Zhiwen/Qwen-Doc.

Autonomous agents capable of perceiving complex environments, understanding instructions, and performing multi-step tasks hold transformative potential across domains such as robotics, scientific discovery, and web automation. While large language models (LLMs) provide a powerful foundation, they struggle with closed-loop decision-making due to static pretraining and limited temporal grounding. Prior approaches either rely on expensive, real-time environment interactions or brittle imitation policies, both with safety and efficiency trade-offs. We introduce DreamPhase, a modular framework that plans through offline imagination. A learned latent world model simulates multi-step futures in latent space; imagined branches are scored with an uncertainty-aware value and filtered by a safety gate. The best branch is distilled into a short natural-language reflection that conditions the next policy query, improving behavior without modifying the LLM. Crucially, DreamPhase attains its performance with substantially fewer real interactions: on WebShop, average API calls per episode drop from $\sim$40 with ARMAP-M (token-level search) to $<10$ with DreamPhase, a $4\times$ reduction that lowers latency and reduces executed irreversible actions by $\sim 5\times$ on WebShop (4.9$\times$ on ALFWorld) per incident logs. Across web, science, and embodied tasks, DreamPhase improves sample efficiency, safety, and cost over search-based and reward-based baselines. This offers a scalable path toward safe, high-performance autonomous agents via imagination-driven planning. Code: \url{https://anonymous.4open.science/r/DreamPhase-A8AD/README.md}.

Achieving generalized in-hand object rotation remains a significant challenge in robotics, largely due to the difficulty of transferring policies from simulation to the real world. The complex, contact-rich dynamics of dexterous manipulation create a "reality gap" that has limited prior work to constrained scenarios involving simple geometries, limited object sizes and aspect ratios, constrained wrist poses, or customized hands. We address this sim-to-real challenge with a novel framework that enables a single policy, trained in simulation, to generalize to a wide variety of objects and conditions in the real world. The core of our method is a joint-wise dynamics model that learns to bridge the reality gap by effectively fitting limited amount of real-world collected data and then adapting the sim policy’s actions accordingly. The model is highly data‑efficient and generalizable across different whole‑hand interaction distributions by factorizing dynamics across joints, compressing system-wide influences into low‑dimensional variables, and learning each joint’s evolution from its own dynamic profile, implicitly capturing these net effects. We pair this with a fully autonomous data collection strategy that gathers diverse, real-world interaction data with minimal human intervention. Our complete pipeline demonstrates unprecedented generality: a single policy successfully rotates challenging objects with complex shapes (*e.g.*, animals), high aspect ratios (up to 5.33), and small sizes, all while handling diverse wrist orientations and rotation axes. Comprehensive real-world evaluations and a teleoperation application for complex tasks validate the effectiveness and robustness of our approach. Website: [DexNDM](https://meowuu7.github.io/DexNDM/).

Quantile Partial Effect (QPE) is a statistic associated with conditional quantile regression, measuring the effect of covariates at different levels. Our theory demonstrates that when the QPE of cause on effect is assumed to lie in a finite linear span, cause and effect are identifiable from their observational distribution. This generalizes previous identifiability results based on Functional Causal Models (FCMs) with additive, heteroscedastic noise, etc. Meanwhile, since QPE resides entirely at the observational level, this parametric assumption does not require considering mechanisms, noise, or even the Markov assumption, but rather directly utilizes the asymmetry of shape characteristics in the observational distribution. By performing basis function tests on the estimated QPE, causal directions can be distinguished, which is empirically shown to be effective in experiments on a large number of bivariate causal discovery datasets. For multivariate causal discovery, leveraging the close connection between QPE and score functions, we find that Fisher Information is sufficient as a statistical measure to determine causal order when assumptions are made about the second moment of QPE. We validate the feasibility of using Fisher Information to identify causal order on multiple synthetic and real-world multivariate causal discovery datasets.

Current video avatar models can generate fluid animations but struggle to capture a character's authentic essence, primarily synchronizing motion with low-level audio cues instead of understanding higher-level semantics like emotion or intent. To bridge this gap, we propose a novel framework for generating character animations that are not only physically plausible but also semantically rich and expressive. Our model is built on two technical innovations. First, we employ Multimodal Large Language Models to generate a structured textual representation from input conditions, providing high-level semantic guidance for creating contextually and emotionally resonant actions. Second, to ensure robust fusion of multimodal signals, we introduce a specialized Multimodal Diffusion Transformer architecture featuring a novel Pseudo Last Frame design. This allows our model to accurately interpret the joint semantics of audio, images and text, generating motions that are deeply coherent with the overall context. Comprehensive experiments validate the superiority of our method, which achieves compelling results in lip-sync accuracy, video quality, motion naturalness, and semantic consistency. The approach also shows strong generalization to challenging scenarios, including multi-person and non-human subjects. Our video results are linked in https://omnihuman-lab.github.io/v1_5/ .

Adversarial training (AT) is an effective defense for large language models (LLMs) against jailbreak attacks, but performing AT on LLMs is costly. To improve the efficiency of AT for LLMs, recent studies propose continuous AT (CAT) that searches for adversarial inputs within the continuous embedding space of LLMs during AT. While CAT has achieved empirical success, its underlying mechanism, i.e., why adversarial perturbations in the embedding space can help LLMs defend against jailbreak prompts synthesized in the input token space, remains unknown. This paper presents the first theoretical analysis of CAT on LLMs based on in-context learning (ICL) theory. For linear transformers trained with adversarial examples from the embedding space on in-context linear regression tasks, we prove a robust generalization bound that has a negative correlation with the perturbation radius in the embedding space. This clearly explains why CAT can defend against jailbreak prompts from the LLM's token space. Further, the robust bound shows that the robustness of an adversarially trained LLM is closely related to the singular values of its embedding matrix. Based on this, we propose to improve LLM CAT by introducing an additional regularization term, which depends on singular values of the LLM's embedding matrix, into the objective function of CAT. Experiments on real-world LLMs demonstrate that our method can help LLMs achieve better jailbreak robustness-utility tradeoff.

Recent advances in 3D Gaussian Splatting (3DGS) enable real-time, high-fidelity novel view synthesis (NVS) with explicit 3D representations. However, performance degradation and instability remain significant under sparse-view conditions. In this work, we identify two key failure modes under sparse-view conditions: overfitting in regions with excessive Gaussian density near the camera, and underfitting in distant areas with insufficient Gaussian coverage. To address these challenges, we propose a unified framework \modelname{}, comprising two key components: a Depth-and-Density Guided Dropout strategy that suppresses overfitting by adaptively masking redundant Gaussians based on density and depth, and a Distance-Aware Fidelity Enhancement module that improves reconstruction quality in under-fitted far-field areas through targeted supervision. Moreover, we introduce a new evaluation metric to quantify the stability of learned Gaussian distributions, providing insights into the robustness of the sparse-view 3DGS. Extensive experiments on multiple datasets demonstrate that our method significantly improves both visual quality and robustness under sparse view conditions. The source code and trained models will be made publicly available.

Electronic medical records (EMRs) are vital for healthcare research, but their use is limited by privacy concerns. Synthetic EMR generation offers a promising alternative, yet most existing methods merely imitate real records without adhering to rigorous clinical quality principles. To address this, we introduce LLM-CARe, a stage-wise cyclic refinement framework that progressively improves EMR quality through three stages, each targeting a specific granularity: corpus, section and document. At each stage, a Critic, an Adviser, and a Reviser collaborate iteratively to evaluate, provide feedback, and refine the drafts. This structured, multi-stage process produces records that better satisfy clinical quality standards. Experiments show that LLM-CARe significantly enhances EMR quality across all levels compared to strong baselines and yields improved performance on real-world clinical tasks such as diagnosis prediction. Unlike prior work, our method requires no real EMR text for training or prompting, demonstrating the effectiveness of stage-wise, cyclic refinement for generating high-quality, privacy-preserving EMR datasets.

Reasoning models have made rapid progress on many benchmarks involving math, code, and science. Yet, there are still many open questions about the best train- ing recipes for reasoning since state-of-the-art models often rely on proprietary datasets with little to no public information available. To address this, the goal of the OpenThoughts project is to create open-source datasets for training reasoning models. Our OpenThoughts2-1M dataset led to OpenThinker2-32B, the first model trained on public reasoning data to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks such as AIME and LiveCodeBench. We then improve our dataset further by systematically investigating each step of our data genera- tion pipeline with 1,000+ controlled experiments, which led to OpenThoughts3. Scaling the pipeline to 1.2M examples and using QwQ-32B as teacher yields our OpenThinker3-7B model, which achieves state-of-the-art results: 53% on AIME 2025, 51% on LiveCodeBench 06/24-01/25, and 54% on GPQA Dia- mond – improvements of 15.3, 17.2, and 20.5 percentage points compared to the DeepSeek-R1-Distill-Qwen-7B. All of our datasets and models are available on ANONYMIZED.

Unified Multimodal Large Language Models (U-MLLMs) have garnered considerable interest for their ability to seamlessly integrate generation and comprehension tasks. However, existing research lacks a unified evaluation standard, often relying on isolated benchmarks to assess these capabilities. Moreover, current work highlights the potential of ``mixed-modality generation capabilities'' through case studies—such as generating auxiliary lines in images to solve geometric problems, or reasoning through a problem before generating a corresponding image. Despite this, there is no standardized benchmark to assess models on such unified tasks. To address this gap, we introduce MME-Unify, also termed as MME-U, the first open and reproducible benchmark designed to evaluate multimodal comprehension, generation, and mixed-modality generation capabilities. For comprehension and generation tasks, we curate a diverse set of tasks from 12 datasets, aligning their formats and metrics to develop a standardized evaluation framework. For unified tasks, we design five subtasks to rigorously assess how models’ understanding and generation capabilities can mutually enhance each other. Evaluation of 12 U-MLLMs, including Janus-Pro, EMU3, and Gemini2-Flash, reveals significant room for improvement, particularly in areas such as instruction following and image generation quality.

The memorization of training data by neural networks raises pressing concerns for privacy and security. Recent work has shown that, under certain conditions, portions of the training set can be reconstructed directly from model parameters. Some of these methods exploit implicit bias toward margin maximization, suggesting that properties often regarded as beneficial for generalization may actually compromise privacy. Yet despite striking empirical demonstrations, the reliability of these attacks remains poorly understood and lacks a solid theoretical foundation. In this work, we take a complementary perspective: rather than designing stronger attacks, we analyze the inherent weaknesses and limitations of existing reconstruction methods and identify conditions under which they fail. We rigorously prove that, without incorporating prior knowledge about the data, there exist infinitely many alternative solutions that may lie arbitrarily far from the true training set, rendering reconstruction fundamentally unreliable. Empirically, we further demonstrate that exact duplication of training examples occurs only by chance. Our results refine the theoretical understanding of when training set leakage is possible and offer new insights into mitigating reconstruction attacks. Remarkably, we demonstrate that networks trained more extensively, and therefore satisfying implicit bias conditions more strongly -- are, in fact, less susceptible to reconstruction attacks, reconciling privacy with the need for strong generalization in this setting.

Discrete diffusion models have become highly effective across various domains. However, real-world applications often require the generative process to adhere to certain constraints. To this end, we propose a Sequential Monte Carlo (SMC) framework that enables scalable inference-time control of discrete diffusion models through principled importance weighting and optimal proposal construction. Specifically, our approach derives tractable importance weights for a range of intermediate targets and characterises the optimal proposal, for which we develop two practical approximations: a first-order gradient-based approximation and an amortised proposal trained to minimise the log-variance of the importance weights. Empirical results across synthetic tasks, language modelling, biology design, and text-to-image generation demonstrate that our framework enhances controllability and sample quality, highlighting the effectiveness of SMC as a versatile recipe for scaling discrete diffusion models at inference time.

Offline reinforcement learning enables policy learning solely from fixed datasets, without costly or risky environment interactions, making it highly valuable for real-world applications. While Transformer-based approaches have recently demonstrated strong sequence modeling capabilities, they typically learn from complete trajectories conditioned on final returns. To mitigate this limitation, we propose the Peak-Return Greedy Slicing (PRGS) framework, which explicitly partitions trajectories at the timestep level and emphasizes high-quality subtrajectories. PRGS first leverages an MMD-based return estimator to characterize the distribution of future returns for state-action pairs, yielding optimistic return estimates. It then performs greedy slicing to extract high-quality subtrajectories for training. During evaluation, an adaptive history truncation mechanism is introduced to align the inference process with the training procedure. Extensive experiments across multiple benchmark datasets indicate that PRGS significantly improves the performance of Transformer-based offline reinforcement learning methods by effectively enhancing their ability to exploit and recombine valuable subtrajectories.

Clinical decision-making is a dynamic, interactive, and cyclic process where doctors have to repeatedly decide on which clinical action to perform and consider newly uncovered information for diagnosis and treatment. Large Language Models (LLMs) have the potential to support clinicians in this process, however, most applications of LLMs in clinical decision support suffer from one of two limitations: Either they assume the unrealistic scenario of immediate availability of all patient information and do not model the interactive and iterative investigation process, or they restrict themselves to the limited "out-of-the-box" capabilities of large pre-trained models without performing task-specific training. In contrast to this, we propose to model clinical decision-making for diagnosis with a hypothesis-driven uncertainty-aware language agent, LA-CDM, that converges towards a diagnosis via repeatedly requesting and interpreting relevant tests. Using a hybrid training paradigm combining supervised and reinforcement learning, we train LA-CDM with three objectives targeting critical aspects of clinical decision-making: accurate hypothesis generation, hypothesis uncertainty estimation, and efficient decision-making. We evaluate our methodology on MIMIC-CDM, a real-world dataset covering four abdominal diseases containing various clinical tests and show the benefit of explicitly training clinical decision-making for increasing diagnostic performance and efficiency. We will release our code upon acceptance.

We propose a post-hoc adaptive conformal anomaly detection method for monitoring time series that leverages predictions from pre-trained foundation models without requiring additional fine-tuning. Our method yields an interpretable anomaly score directly interpretable as a false alarm rate (p-value), facilitating transparent and actionable decision-making. It employs weighted quantile conformal prediction bounds and adaptively learns optimal weighting parameters from past predictions, enabling calibration under distribution shifts and stable false alarm control, while preserving out-of-sample guarantees. As a model-agnostic solution, it integrates seamlessly with foundation models and supports rapid deployment in resource-constrained environments. This approach addresses key industrial challenges such as limited data availability, lack of training expertise, and the need for immediate inference, while taking advantage of the growing accessibility of time series foundation models. Experiments on both synthetic and real-world datasets show that the proposed approach delivers strong performance, combining simplicity, interpretability, robustness, and adaptivity.

Vision Language Models (VLMs) show strong potential for visual planning but struggle with precise spatial and long-horizon reasoning. In contrast, Planning Domain Definition Language (PDDL) planners excel at long-horizon formal planning, but cannot interpret visual inputs. Recent works combine these complementary advantages by enabling VLMs to turn visual planning problems into PDDL files for formal planning. However, while VLMs can generate PDDL problem files satisfactorily, they struggle to accurately generate the PDDL domain files, which describe all the planning rules. As a result, prior methods rely on human experts to predefine domain files or on constant environment access for refinement. We propose VLMFP, a Dual-VLM-guided framework that can autonomously generate both PDDL problem and domain files for formal visual planning. VLMFP introduces two VLMs to ensure reliable PDDL file generation: A SimVLM that simulates action consequences based on input rule descriptions, and a GenVLM that generates and iteratively refines PDDL files by comparing the PDDL and SimVLM execution results. VLMFP unleashes multiple levels of generalizability: The same generated PDDL domain file works for all the different instances under the same problem, and VLMs generalize to different problems with varied appearances and rules. We evaluate VLMFP with 6 grid-world domains and test its generalization to unseen instances, appearance, and game rules. On average, SimVLM accurately describes 95.5\%, 82.6\% of scenarios, simulates 85.5\%, 87.8\% of action sequence, and judges 82.4\%, 85.6\% goal reaching for seen and unseen appearances, respectively. With the guidance of SimVLM, VLMFP can generate PDDL files to reach 70.0\%, 54.1\% valid plans for unseen instances in seen and unseen appearances, respectively.

We introduce GradPCA, an Out-of-Distribution (OOD) detection method that exploits the low-rank structure of neural network gradients induced by Neural Tangent Kernel (NTK) alignment. GradPCA applies Principal Component Analysis (PCA) to gradient class-means, achieving more consistent performance than existing methods across standard image classification benchmarks. We provide a theoretical perspective on spectral OOD detection in neural networks to support GradPCA, highlighting feature-space properties that enable effective detection and naturally emerge from NTK alignment. Our analysis further reveals that feature quality—particularly the use of pretrained versus non-pretrained representations—plays a crucial role in determining which detectors will succeed. Extensive experiments validate the strong performance of GradPCA, and our theoretical framework offers guidance for designing more principled spectral OOD detectors.

The performance of reinforcement learning (RL) agents depends critically on the quality of the underlying feature representations. Hyperbolic feature spaces are well-suited for this purpose, as they naturally capture hierarchical and relational structure often present in complex RL environments. However, leveraging these spaces commonly faces optimization challenges due to the nonstationarity of RL. In this work, we identify key factors that determine the success and failure of training hyperbolic deep RL agents. By analyzing the gradients of core operations in the Poincaré ball and hyperboloid models of hyperbolic geometry, we show that large-norm embeddings destabilize gradient-based training, leading to trust-region violations in proximal policy optimization (PPO). Based on these insights, we introduce Hyper++, a new hyperbolic PPO agent that consists of three components: (1) stable critic training through a categorical value loss instead of regression; (2) feature regularization guaranteeing bounded norms while avoiding the curse of dimensionality from clipping; and (3) using a more optimization-friendly formulation of hyperbolic network layers. In experiments on ProcGen, we show that Hyper++ guarantees stable learning, outperforms prior hyperbolic agents, and reduces wall-clock time by approximately 30%. On Atari-5 with Double DQN, Hyper++ strongly outperforms Euclidean and hyperbolic baselines.