ICLR 2026 Papers

Artificial Intelligence Generated Content (AIGC), particularly video generation with diffusion models, has been advanced rapidly. Invisible watermarking is a key technology for protecting AI-generated videos and tracing harmful content, and thus plays a crucial role in AI safety. Beyond post-processing watermarks which inevitably degrade video quality, recent studies have proposed distortion-free in-generation watermarking for video diffusion models. However, existing in-generation approaches are non-blind: they require maintaining all the message-key pairs and performing template-based matching during extraction, which incurs prohibitive computational costs at scale. Moreover, when applied to modern video diffusion models with causal 3D Variational Autoencoders (VAEs), their robustness against temporal disturbance becomes extremely weak. To overcome these challenges, we propose SIGMark, a Scalable In-Generation watermarking framework with blind extraction for video diffusion. To achieve blind-extraction, we propose to generate watermarked initial noise using a Global set of Frame-wise PseudoRandom Coding keys (GF-PRC), reducing the cost of storing large-scale information while preserving noise distribution and diversity for distortion-free watermarking. To enhance robustness, we further design a Segment Group-Ordering module (SGO) tailored to causal 3D VAEs, ensuring robust watermark inversion during extraction under temporal disturbance. Comprehensive experiments on modern diffusion models show that SIGMark achieves very high bit-accuracy during extraction under both temporal and spatial disturbances with minimal overhead, demonstrating its scalability and robustness.

Long-context models are essential for many applications but face inefficiencies in loading large KV caches during decoding. Prior methods enforce fixed token budgets for sparse attention, assuming a set number of tokens can approximate full attention. However, these methods overlook variations in the importance of attention across heads, layers, and contexts. To address these limitations, we propose Tactic, a sparsity-adaptive and calibration-free sparse attention mechanism that dynamically selects tokens based on their cumulative attention scores rather than a fixed token budget. By setting a target fraction of total attention scores, Tactic ensures that token selection naturally adapts to variations in attention sparsity. To efficiently approximate this selection, Tactic leverages clustering-based sorting and distribution fitting, allowing it to accurately estimate token importance with minimal computational overhead. We show that Tactic outperforms existing sparse attention algorithms, achieving superior accuracy and up to 5.14x decode attention speedup. This improvement translates to an overall 1.51x end-to-end inference speedup, making Tactic a practical and effective solution for long-context LLM inference in accuracy-sensitive applications.

The ability to generate virtual environments is crucial for applications ranging from gaming to physical AI domains such as robotics, autonomous driving, and industrial AI. Current learning-based 3D reconstruction methods rely on the availability of captured real-world multi-view data, which is not always readily available. Recent advancements in video diffusion models have shown remarkable imagination capabilities, yet their 2D nature prevents their use in simulations where a robot needs to navigate and interact with the environment. In this paper, we propose a self-distillation framework that aims to distill the implicit 3D knowledge in the video diffusion models into an explicit 3D Gaussian Splatting (3DGS) representation, eliminating the need for multi-view training data. Specifically, we augment the typical RGB decoder with a 3DGS decoder, which is supervised by the output of the RGB decoder. In this approach, the 3DGS decoder can be purely trained with synthetic data generated by video diffusion models. At inference time, our model can synthesize 3D scenes from either a text prompt or a single image for real-time rendering. Our framework further extends to dynamic 3D scene generation from a monocular input video. Experimental results show that our framework achieves state-of-the-art performance in static and dynamic 3D scene generation. Video results: https://research.nvidia.com/labs/toronto-ai/lyra

Heuristic design with large language models (LLMs) has emerged as a promising approach for tackling combinatorial optimization problems (COPs). However, existing approaches often rely on manually predefined evolutionary computation (EC) heuristic-optimizers and single-task training schemes, which may constrain the exploration of diverse heuristic algorithms and hinder the generalization of the resulting heuristics. To address these issues, we propose Meta-Optimization of Heuristics (MoH), a novel framework that operates at the optimizer level, discovering effective heuristic-optimizers through the principle of meta-learning. Specifically, MoH leverages LLMs to iteratively refine a meta-optimizer that autonomously constructs diverse heuristic-optimizers through (self-)invocation, thereby eliminating the reliance on a predefined EC heuristic-optimizer. These constructed heuristic-optimizers subsequently evolve heuristics for downstream tasks, enabling broader heuristic exploration. Moreover, MoH employs a multi-task training scheme to promote its generalization capability. Experiments on classic COPs demonstrate that MoH constructs an effective and interpretable meta-optimizer, achieving state-of-the-art performance across various downstream tasks, particularly in cross-size settings.

Representation learning is critical to the empirical and theoretical success of reinforcement learning. However, many existing methods are induced from model-learning aspects, misaligning them with the RL task in hand. This work introduces the Spectral Bellman Method, a novel framework derived from the Inherent Bellman Error (IBE) condition. It aligns representation learning with the fundamental structure of Bellman updates across a space of possible value functions, making it directly suited for value-based RL. Our key insight is a fundamental spectral relationship: under the zero-IBE condition, the transformation of a distribution of value functions by the Bellman operator is intrinsically linked to the feature covariance structure. This connection yields a new, theoretically-grounded objective for learning state-action features that capture this Bellman-aligned covariance, requiring only a simple modification to existing algorithms. We demonstrate that our learned representations enable structured exploration by aligning feature covariance with Bellman dynamics, improving performance in hard-exploration and long-horizon tasks. Our framework naturally extends to multi-step Bellman operators, offering a principled path toward learning more powerful and structurally sound representations for value-based RL.

Large language models (LLMs) with Process Reward Models (PRMs) have shown strong reasoning ability, yet their potential in Operations Research (OR) remains unexplored. We present the first PRM tailored for OR, but find that directly training on mainstream datasets yields surprisingly weak performance. To understand this gap, we conduct a systematic analysis and identify the primary bottleneck: the datasets themselves, where over 30\% of annotations are severely flawed. To overcome these limitations, we first collect all existing synthetic datasets and apply a carefully designed filtering pipeline to construct a high-quality seed dataset. Building upon this seed, we then build OR-ProcessQA, the first large-scale dataset for OR with step-by-step supervision, where diverse solution pathways are generated via Monte Carlo Tree Search (MCTS) and each step is validated for logical consistency by GPT-4o. Building on this foundation, we train OR-PRM, the first Process Reward Model in the OR domain, designed to evaluate and guide reasoning at every step rather than only the final outcome. Together, these advances enable OR-PRM to substantially improve LLMs’ reasoning capability, achieving a maximum absolute improvement of 12.5\% over the base model in Best-of-N settings, and highlighting the power of process-oriented supervision for reliable problem solving in operations research.

We present a deep learning model for high-resolution probabilistic precipitation forecasting over an 8-hour horizon in Europe, overcoming the limitations of radar-only deep learning models with short forecast lead times. Our model efficiently integrates multiple data sources - including radar, satellite, and physics-based numerical weather prediction (NWP) - while capturing long-range interactions, resulting in accurate forecasts with robust uncertainty quantification through consistent probabilistic maps. Featuring a compact architecture, it enables more efficient training and faster inference than existing models. Extensive experiments demonstrate that our model surpasses current operational NWP systems, extrapolation-based methods, and deep-learning nowcasting models, setting a new standard for high-resolution precipitation forecasting in Europe, ensuring a balance between accuracy, interpretability, and computational efficiency.

We introduce a new Reflective Generative Model (RGM), which obtains OpenAI o3-mini's performance via a novel Reflective Generative Form. This form focuses on high-quality reasoning trajectory selection and contains two novelties: 1) A unified interface for policy and process reward model: we share the backbone network and use task-specific heads for reasoning trajectory predicting and scoring respectively, introducing only 50M extra parameters for trajectory scoring. 2) Eliminating the reliance on process-level annotation: we provide a self-supervised process reward model (SPRM), which can directly learn the high-quality reasoning trajectory selection from the outcome reward. Equipped with the reflective generative form, RGM is naturally suitable for test-time scaling based on the controllable thinking length. Experiments show that our RGM, equipped with only 50M additional parameters in SPRM, outperforms policy models with 72B extra reward models, thereby enabling QwQ-32B to outperform OpenAI o3-mini on AIME24 (84.2 vs. 79.6) and HMMT25 (53.1 vs. 53.0). Code will be available.

Humans acquire semantic object representations from egocentric visual streams with minimal supervision. Importantly, the visual system processes with high resolution only the center of its field of view and learns similar representations for visual inputs occurring close in time. This emphasizes slowly changing information around gaze locations. This study investigates the role of central vision and slowness learning in the formation of semantic object representations in humans. We simulate five months of human-like visual experience using the Ego4D dataset and generate gaze coordinates with a state-of-the-art gaze prediction model. Using these predictions, we extract crops that mimic central vision and train a time-contrastive Self-Supervised Learning model on them. Our results show that combining temporal slowness and central vision improves the encoding of different semantic facets of object representations. Specifically, focusing on central vision strengthens the extraction of foreground object features, while considering temporal slowness, especially during fixational eye movements, allows the model to encode broader semantic information about objects. These findings provide new insights into the mechanisms by which humans may develop semantic object representations from natural visual experience. Our code will be made public upon acceptance.

A simple yet effective method for inference-time alignment of generative models is Best-of-$N$ (BoN), where $N$ outcomes are sampled from a reference policy, evaluated using a proxy-reward model, and the highest-scoring one is selected. While prior work argues that BoN is almost optimal in reward vs KL tradeoffs, the effectiveness of BoN depends critically on the quality of the proxy-reward model used for selection. For this purpose, we study BoN through a smooth version known as Soft Best-of-N (SBoN) and develop a theoretical framework to address this gap. We analyze the scaling behaviour of BoN by providing bounds on the KL divergence between the SBoN policy and the reference policy, offering insights into how performance varies with the number of samples. We also study the regret gap, i.e., the gap between the expected true-reward under the optimal policy and the SBoN policy. Our theoretical and empirical findings show that smoothing helps SBoN mitigate reward overoptimization, especially when the quality of the proxy-reward is low.

The discovery of novel catalytic materials is a cornerstone of chemical engineering and sustainable energy, yet it remains a complex, knowledge-intensive process. While Large Language Models (LLMs) have demonstrated remarkable potential in various scientific domains, their application to catalysis is hindered by the lack of specialized, multi-dimensional benchmarks to guide their development and evaluation. To bridge the critical gap, we introduce CatalystBench, a comprehensive and challenging benchmark meticulously constructed from scientific literature and public datasets, specifically designed to assess the capabilities of LLMs in the nuanced domain of catalyst design. The tasks covered by this benchmark dataset encompass the entire closed-loop process of catalyst development, including reading comprehension, experimental analysis, and scheme reasoning. Based on this benchmark, we propose a Multi-head Full-task (MFT) domain-specific fine-tuning method that employs coupling task-specific output heads. We systematically compare with other three distinct fine-tuning strategies: Single-Task (ST), Full-Task (FT) and Multi-head Single-Task (MST). The extensive experiments demonstrate that the MFT strategy consistently achieves the most substantial performance improvements across all tasks, underscoring the effectiveness of explicit multi-task architectures in complex scientific reasoning. The resulting CatalystLLM significantly outperforms a wide array of state-of-the-art open-source and closed-source models on CatalystBench. We will publicly release both the CatalystBench benchmark and the CatalystLLM model, providing the community with a robust evaluation framework and a powerful new tool to accelerate AI-driven research in catalytic materials science.

Recent studies have demonstrated that test-time compute scaling effectively improves the performance of small language models (sLMs). However, prior research has mainly examined test-time compute scaling with an additional larger model as a verifier, leaving verification by sLMs underexplored. In this work, we investigate whether sLMs can reliably verify the output candidates under test-time scaling. We find that even with knowledge distillation from larger verifiers, sLMs struggle with verification tasks requiring memorization, such as numerical calculations and fact-checking. To address this limitation, we propose Tool-integrated verification (T1), a two-stage framework that first filters candidates with external tools and then uses an sLM for final verification, offloading memorization-heavy steps to tools such as a code interpreter. Within T1 we prove that offloading to external tools reduces the memorization burden on sLMs and improves test-time scaling performance. Experiments on the MATH benchmark demonstrate that, with T1, a Llama-3.2 1B model under test-time scaling outperforms the significantly larger Llama-3.1 8B model. Moreover, T1 improves the verification accuracy of both process reward models (PRMs) and critic models. Our findings highlight the potential of tool integration to substantially improve the verification abilities of sLMs.

Reinforcement learning based post-training has recently emerged as a powerful paradigm for enhancing the alignment and reasoning capabilities of multimodal large language models (MLLMs). While *vision-centric* post-training is crucial for enhancing MLLMs’ intrinsic understanding of visual signals, current post-training paradigms are predominantly *text-centric*, where dense visual inputs are only leveraged to extract sparse cues for text-based reasoning. There exist a few approaches in this direction, however, they often still rely on text as an intermediate mediator or introduce additional visual generative designs. In this work, we introduce **Visual Jigsaw**, a generic *self-supervised* post-training framework designed to strengthen visual understanding in MLLMs. Visual Jigsaw is formulated as a general ordering task: visual inputs are partitioned, shuffled, and the model must reconstruct the visual information by producing the correct permutation in natural language. This naturally aligns with reinforcement learning from verifiable rewards (RLVR), requires no additional visual generative components, and derives its supervisory signal automatically without any annotations. We instantiate Visual Jigsaw across three visual modalities, including images, videos, and 3D data. Extensive experiments demonstrate substantial improvements in fine-grained perception, temporal reasoning, and 3D spatial understanding. Our findings highlight the potential of self-supervised vision-centric tasks in post-training MLLMs and aim to inspire further research on vision-centric pretext designs.

Referring Image Segmentation (RIS) aims to segment the object in an image uniquely referred to by a natural language expression. However, RIS training often contains hard-to-align and instance-specific visual signals; optimizing on such pixels injects misleading gradients and drives the model in the wrong direction. By explicitly estimating pixel-level vision–language alignment, the learner can suppress low-alignment regions, concentrate on reliable cues, and acquire more generalizable alignment features. In this paper, we propose Alignment-Aware Masked Learning (AML), a simple yet effective training strategy that quantifies region–referent alignment (PMME) and filters out unreliable pixels during optimization (AFM). Specifically, each sample first computes a similarity map between visual and textual features, and then masks out pixels falling below an adaptive similarity threshold, thereby excluding poorly aligned regions from the training process. AML does not require architectural changes and incurs no inference overhead, directing attention to the areas aligned with the textual description. Experiments on the RefCOCO (vanilla/+/g) datasets show that AML achieves state-of-the-art results across all 8 splits, and beyond improving RIS performance, AML also enhances the model’s robustness to diverse descriptions and scenarios.

Physics-constrained generative modeling aims to produce high-dimensional samples that are both physically consistent and distributionally accurate, a task that remains challenging due to often conflicting optimization objectives. Recent advances in flow matching and diffusion models have enabled efficient generative modeling, but integrating physical constraints often degrades generative fidelity or requires costly inference-time corrections. Our work is the first to recognize the trade-off between distributional and physical accuracy. Based on the insight of inherently conflicting objectives, we introduce Physics-Based Flow Matching (PBFM) a method that enforces physical constraints at training time using conflict-free gradient updates and unrolling to mitigate Jensen's gap. Our approach avoids manual loss balancing and enables simultaneous optimization of generative and physical objectives. As a consequence, physics constraints do not impede inference performance. We benchmark our method across three representative PDE benchmarks. PBFM achieves a Pareto-optimal trade-off, competitive inference speed, and generalizes to a wide range of physics-constrained generative tasks, providing a practical tool for scientific machine learning. Code and datasets available at [https://github.com/tum-pbs/PBFM](https://github.com/tum-pbs/PBFM).

Biological profiling technologies, such as imaging mass cytometry (IMC) and spatial transcriptomics (ST), generate multi-channel data with strong spatial alignment and complex inter-channel relationships. Modeling such data requires generative frameworks that can jointly model spatial structure and channel relationships, while also generalizing across arbitrary combinations of observed and missing channels for practical applications. Existing generative models typically assume low-dimensional inputs (e.g., RGB images) and rely on simple conditioning mechanisms that break spatial correspondence and overlook inter-channel dependencies. This work proposes a unified multi-channel diffusion (MCD) framework for controllable generation of structured biological data with intricate inter-channel relationships. Our model introduces two key innovations: (1) a hierarchical feature injection mechanism that enables multi-resolution conditioning on spatially aligned observed channels, and (2) two complementary channel attention modules to capture inter-channel relationships and recalibrate latent features. To support flexible conditioning and generalization to arbitrary sets of observed channels, we train the model using a random channel masking strategy, enabling it to reconstruct missing channels from any combination of observed channels as the spatial condition. We demonstrate state-of-the-art performance across both spatial and non-spatial biological data generation tasks, including imputation in spatial proteomics and clinical imaging, as well as gene-to-protein prediction in single-cell datasets, and show strong generalizability to unseen conditional configurations.

In this paper, we propose a rotation-constrained compensation method to address the errors introduced by structured pruning of large language models (LLMs). LLMs are trained on massive datasets and accumulate rich semantic knowledge in their representation space. In contrast, pruning is typically carried out with only a small amount of calibration data, which makes output mismatches unavoidable. Although direct least-squares fitting can reduce such errors, it tends to overfit to the limited calibration set, destructively modifying pretrained weights. To overcome this difficulty, we update the pruned parameters under a rotation constraint. This constrained update preserves the geometry of output representations (i.e., norms and inner products) and simultaneously re-aligns the pruned subspace with the original outputs. Furthermore, in rotation-constrained compensation, removing components that strongly contribute to the principal directions of the output makes error recovery difficult. Since input dimensions with large variance strongly affect these principal directions, we design a variance-aware importance score that ensures such dimensions are preferentially kept in the pruned model. By combining this scoring rule with rotation-constrained updates, the proposed method effectively compensates errors while retaining the components likely to be more important in a geometry-preserving manner. In the experiments, we apply the proposed method to Llama-7B and Llama-2-13B, and evaluate it on WikiText-2 and multiple language understanding benchmarks. The results demonstrate consistently better perplexity and task accuracy compared with existing baselines. Codes are available at \url{https://anonymous.4open.science/r/anonymous-llm-pruning-D884/}.

Recent progress in in-context reinforcement learning (ICRL) has demonstrated its potential for training generalist agents that can acquire new tasks directly at inference. Algorithm Distillation (AD) pioneered this paradigm and was subsequently scaled to multi-domain settings, although its ability to generalize to unseen tasks remained limited. The Decision Pre-Trained Transformer (DPT) was introduced as an alternative, showing stronger in-context reinforcement learning abilities in simplified domains, but its scalability had not been established. In this work, we extend DPT to diverse multi-domain environments, applying Flow Matching as a natural training choice that preserves its interpretation as Bayesian posterior sampling. As a result, we obtain an agent trained across hundreds of diverse tasks that achieves clear gains in generalization to the held-out test set. This agent improves upon prior AD scaling and demonstrates stronger performance in both online and offline inference, reinforcing ICRL as a viable alternative to expert distillation for training generalist agents.

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs for text generation, with the potential to decode multiple tokens in a single iteration. However, none of the existing open-source dLLMs have achieved superior inference speed over AR LLMs of similar size. This paper breaks this barrier based on a simple and effective strategy named discrete diffusion forcing (D2F). D2F equips dLLMs with two key capabilities: (1) block-wise autoregressive generation to enable KV cache utilization; (2) prediction of following tokens without requiring completion of prior blocks for inter-block parallel decoding. In this way, the vanilla dLLMs are refurbished into an AR-diffusion hybrid paradigm for efficient inference. D2F can be implemented with an asymmetric distillation process based on pre-trained dLLMs to achieve rapid convergence.We further propose a pipelined parallel decoding algorithm, which enables a trade-off between efficiency and efficacy. Empirically, D2F dLLMs achieve more than $\mathbf{2.5\times}$ inference speed than LLaMA3 and Qwen2.5 on GSM8K. Compared to the vanilla dLLMs like LLaDA and Dream, the acceleration can be more than $\mathbf{50\times}$ while maintaining comparable output quality.