ICLR 2026 Papers

Vision-language-action models (VLAs) trained on large-scale robotic datasets have demonstrated strong performance on manipulation tasks, including bimanual tasks. However, because most public datasets focus on single-arm demonstrations, adapting VLAs for bimanual tasks typically requires substantial additional bimanual data and fine-tuning. To address this challenge, we introduce TwinVLA, a modular framework that composes two copies of a pretrained single-arm VLA into a coordinated bimanual VLA. Unlike monolithic cross-embodiment models trained on mixtures of single-arm and bimanual data, TwinVLA improves both data efficiency and performance by composing pretrained single-arm policies. Across diverse bimanual tasks in real-world and simulation settings, TwinVLA outperforms a comparably-sized monolithic RDT-1B model without requiring any bimanual pretraining. Furthermore, it narrows the gap to state-of-the-art model, $\pi_0$ which rely on extensive proprietary bimanual data and compute cost. These results establish our modular composition approach as a data-efficient and scalable path toward high-performance bimanual manipulation, leveraging public single-arm data.

Degradation-agnostic image restoration aims to handle diverse corruptions with one unified model, but faces fundamental challenges in balancing efficiency and performance across different degradation types. Existing approaches either sacrifice efficiency for versatility or fail to capture the distinct representational requirements of various degradations. We present MIRAGE, an efficient framework that addresses these challenges through two key innovations. First, we propose a channel-wise functional decomposition that systematically repurposes channel redundancy in attention mechanisms by assigning CNN, attention, and MLP branches to handle local textures, global context, and channel statistics, respectively. This principled decomposition enables degradation-agnostic learning while achieving superior efficiency-performance trade-offs. Second, we introduce manifold regularization that performs cross-layer contrastive alignment in Symmetric Positive Definite (SPD) space, which empirically improves feature consistency and generalization across degradation types. Extensive experiments across five degradation settings demonstrate that MIRAGE achieves state-of-the-art performance with remarkable efficiency, outperforming existing methods in both single and mixed degradation scenarios while showing strong zero-shot generalization to unseen domains.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a strong assumption that the collected preference data is clean and that all observed labels are equally reliable. In reality, large-scale preference datasets contain substantial label noise due to annotator errors, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Robust Preference Optimization (RPO). RPO employs an Expectation-Maximization algorithm to infer the posterior probability of each label’s correctness, which is used to adaptively re-weigh each data point in the training loss to mitigate noise. We further generalize this approach by establishing a theoretical link between arbitrary preference losses and their corresponding probabilistic models. This generalization enables the systematic transformation of existing alignment algorithms into their robust counterparts, elevating RPO from a specific algorithm to a general framework for robust preference alignment. Theoretically, we prove that under the condition of a perfectly calibrated model, RPO is guaranteed to converge to the true noise level of the dataset. Our experiments demonstrate RPO’s effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO). When applied to Mistral and Llama 3 models, the RPO-enhanced methods improve AlpacaEval 2 win rates by up to 7.0 percentage points over their respective baselines.

The success of large pretrained Transformers is closely tied to tokenizers, which convert raw input into discrete symbols. Extending these models to graph-structured data remains a significant challenge. In this work, we introduce a graph tokenization framework that generates sequential representations of graphs by combining reversible graph serialization, which preserves graph information, with Byte Pair Encoding (BPE), a widely adopted tokenizer in large language models (LLMs). To better capture structural information, the graph serialization process is guided by global statistics of graph substructures, ensuring that frequently occurring substructures appear more often in the sequence and can be merged by BPE into meaningful tokens. Empirical results demonstrate that the proposed tokenizer enables Transformers such as BERT to be directly applied to graph benchmarks without architectural modifications. The proposed approach achieves state-of-the-art results on 12 benchmark datasets and frequently outperforms both graph neural networks and specialized graph transformers. This work bridges the gap between graph-structured data and the ecosystem of sequence models.

Multimodal learning is of continued interest in artificial intelligence-based applications, motivated by the potential information gain from combining different types of data. However, modalities observed in the source environment may differ from the modalities observed in the target environment due to multiple factors, including cost, hardware failure, or the perceived *informativeness* of a given modality. This shift in missingness between the source and target environment has not been carefully studied. Naive estimation of the information gain associated with including an additional modality without accounting for missingness may result in improper estimates of that modality's value in the target environment. We formalize the problem of missingness, demonstrate its ubiquity, and show that the subsequent distribution shift results in bias when the missingness process is not explicitly accounted for. To address this issue, we introduce ICYM2I (In Case You Multimodal Missed It), a framework for the evaluation of predictive performance and information gain under missingness through inverse probability weighting-based correction. We demonstrate the importance of the proposed adjustment to estimate information gain under missingness on synthetic, semi-synthetic, and real-world datasets.

Diffusion model distillation has emerged as a powerful technique for creating efficient few-step and single-step generators. Among these, Distribution Matching Distillation (DMD) and its variants stand out for their impressive performance, which is widely attributed to their core mechanism of matching the student's output distribution to that of a pre-trained teacher model. In this work, we challenge this conventional understanding. Through a rigorous decomposition of the DMD training objective, we reveal that the primary driver of few-step generation is not the distribution matching term, but a previously overlooked component we identify as \textit{\textbf{C}FG \textbf{A}ugmentation} (\textbf{CA}). We demonstrate that this term acts as the core "engine" of distillation, while the \textbf{D}istribution \textbf{M}atching (\textbf{DM}) term functions as a "regularizer" that ensures training stability and mitigates artifacts. We further validate this decoupling by demonstrating that while the DM term is a highly effective regularizer, it is not unique; simpler non-parametric constraints or GAN-based objectives can serve the same stabilizing function, albeit with different trade-offs. This decoupling of labor between CA and DM also allows a more principled analysis of the properties of both terms, leading to a more systematic and in-depth understanding. This new understanding enables us to propose principled modifications to the distillation process, such as decoupling the noise schedules for the engine and the regularizer, leading to further performance gains.

Test-time domain generalization (TTDG) methods enhance the performance of neural networks on target domains by transferring the feature distribution of target samples to approximate that of the source domain, while avoiding the computational cost associated with fine-tuning on the target domain. However, existing TTDG methods primarily rely on style transfer strategies operating at a coarse granularity, which prove ineffective for pixel-level prediction tasks such as image super-resolution (SR). To address this limitation, we propose a multi-codebook based test-time domain generalization framework (MC-TTDG). Our method leverages both domain-specific and domain-invariant codebooks to achieve fine-grained representation learning on source domains, and performs pixel-level nearest-neighbor feature matching and transfer to accurately adjust target domain features. Furthermore, we introduce a voting-based strategy for optimal domain-specific codebook selection, which improves the precision of feature transfer through multi-party consensus. Extensive experiments across diverse data distributions, and network architectures demonstrate that the proposed method effectively transfers feature distributions for SR networks. Our code is available at https://github.com/ZaizuoTang/MC-TTDG.

Existing Image Manipulation Localization (IML) methods rely heavily on task-specific designs, making them perform well only on the target IML task, while joint training on multiple IML tasks causes significant performance degradation, hindering real applications. To this end, we propose Omni-IML, the first generalist model designed to unify IML across diverse tasks. Specifically, Omni-IML achieves generalization through three key components: (1) a Modal Gate Encoder, which adaptively selects the optimal encoding modality per sample, (2) a Dynamic Weight Decoder, which dynamically adjusts decoder filters to the task at hand, and (3) an Anomaly Enhancement module that leverages box supervision to highlight the tampered regions and facilitate the learning of task-agnostic features. Beyond localization, to support interpretation of the tampered images, we construct Omni-273k, a large high-quality dataset that includes natural language descriptions of tampered artifacts. It is annotated through our automatic, chain-of-thoughts annotation technique. We also design a simple-yet-effective interpretation module to better utilize these descriptive annotations. Our extensive experiments show that our single Omni-IML model achieves state-of-the-art performance across all four major IML tasks, providing a valuable solution for practical deployment and a promising direction of generalist models in image forensics. We will release our code and dataset.

The Successor Measure (SM), a powerful method in reinforcement learning (RL), describes discounted future state distributions under a policy, and it has recently been studied using generative modeling techniques. Although SM is a powerful predictive object, it lacks compact representations tailored for online RL. To address this, we introduce Successor Flow Features (SF2), a representation learning framework that bridges SM estimation with policy optimization. SF2 leverages flow-matching generative models to approximate successor measures, while enforcing a structured linear decomposition into a time-invariant embedding and a time-dependent projection. This yields compact, policy-aware state-action features that integrate readily into standard off-policy algorithms like TD3 and SAC. Experiments on DeepMind Control Suite tasks show that SF2 improves sample efficiency and training stability compared to strong successor feature baselines. We attribute these gains to the compact representation induced by flow matching, which reduces compounding errors in long-horizon predictions.

Recent research has generated hope that inference scaling, such as resampling solutions until they pass verifiers like unit tests, could allow weaker models to match stronger ones. Beyond inference, this approach also enables training reasoning models, where data is curated using rejection sampling against a verifier. However, we show that this approach is fundamentally limited when verifiers are imperfect and have a non-zero probability of producing false positives. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling, regardless of compute budget. Our analysis shows that there is a strong correlation between the model’s single-sample accuracy and its false positive rate on HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model. Empirical results show that optimal sampling attempts are often fewer than 10, as the negative utility of false positives outweighs benefits, bending inference scaling curves downward. Finally, false positives may have other undesirable qualities, like poor adherence to coding style conventions.

Multimodal Large Language Models (MLLMs) have achieved notable gains in various tasks by incorporating Chain-of-Thought (CoT) reasoning in language spaces. Recent work extends this direction by leveraging external tools for visual editing, thereby enhancing the visual signal along the reasoning trajectories. Nevertheless, these approaches remain fundamentally constrained: reasoning is still confined to the language space, with visual information treated as static preconditions. We introduce Latent Visual Reasoning (LVR), a new paradigm that enables autoregressive reasoning directly in the visual embedding space. A visual encoder first projects images into visual tokens within a joint semantic space shared with the language model. The language model is then trained to generate latent states that reconstruct key visual tokens critical for answering the query, constituting the process of latent visual reasoning. By interleaving LVR with standard text generation, our model achieves substantial gains on perception-intensive visual question answering tasks. In addition, we adapt the GRPO algorithm to conduct reinforcement learning on latent reasoning, further balancing LVR and textual generation. We show that LVR substantially improves fine-grained visual understanding and perception, achieving 71.67\% on MMVP compared to 66.67\% with Qwen2.5-VL. Code base and model weights will be released later.

We introduce TINKER, a novel framework for high-fidelity 3D editing without any per-scene finetuning, where only a single edited image (one-shot) or a few edited images (few-shot) are required as input. Unlike prior techniques that demand extensive per-scene optimization to ensure multi-view consistency or to produce dozens of consistent edited input views, TINKER delivers robust, multi-view consistent edits from as few as one or two images. This capability stems from repurposing pretrained diffusion models, which unlocks their latent 3D awareness. To drive research in this space, we curate the first large-scale multi-view editing dataset and data pipeline, spanning diverse scenes and styles. Building on this dataset, we develop our framework capable of generating multi-view consistent edited views without per-scene training, which consists of two novel components: (1) Multi-view consistent editor: Enables precise, reference-driven edits that remain coherent across all viewpoints. (2) Any-view-to-video scene completion model : Leverages spatial-temporal priors from video diffusion to perform high-quality scene completion and novel-view generation even from sparse inputs. Through extensive experiments, TINKER significantly reduces the barrier to generalizable 3D content creation, achieving state-of-the-art performance on editing, novel-view synthesis, and rendering enhancement tasks, while also demonstrating strong potential for 4D editing. We believe that TINKER represents a key step towards truly scalable, zero-shot 3D and 4D editing.

Large-scale pre-training has fundamentally changed how machine learning research is done today: large foundation models are trained once, and then can be used by anyone in the community (including those without data or compute resources to train a model from scratch) to adapt and fine-tune to specific tasks. Applying this same framework to reinforcement learning (RL) is appealing because it offers compelling avenues for addressing core challenges in RL, including sample efficiency and robustness. However, there remains a fundamental challenge to pre-train large models in the context of RL: actions have long-term dependencies, so training a foundation model that reasons across *time* is important. Recent advances in generative AI have provided new tools for modeling highly complex distributions. In this paper, we build a probabilistic model to predict which states an agent will visit in the temporally distant future (i.e., an occupancy measure) using flow matching. As large datasets are often constructed by many distinct users performing distinct tasks, we include in our model a latent variable capturing the user intention. This intention increases the expressivity of our model, and enables adaptation with generalized policy improvement. We call our proposed method **intention-conditioned flow occupancy models (InFOM)**. Comparing with alternative methods for pre-training, our experiments on $36$ state-based and $4$ image-based benchmark tasks demonstrate that the proposed method achieves $1.8 \times$ median improvement in returns and increases success rates by $36\\%$.

When an LLM learns a new fact during finetuning (e.g., new movie releases, updated celebrity gossip, etc.), where does this information go? Are entities enriched with relation information, or do models recall information just-in-time before a prediction? Are ``all of the above'' true with LLMs implementing multiple redundant heuristics? Existing localization approaches (e.g., activation patching) are ill-suited for this analysis because they usually replace parts of the residual stream, thus overriding previous information. To fill this gap, we propose dynamic weight grafting, a technique that selectively grafts weights from a finetuned model onto a pretrained model. Using this technique, we show two separate pathways for retrieving finetuned relation information: 1) "enriching" the residual stream with relation information while processing the tokens that correspond to an entity (e.g., "Zendaya" in "Zendaya co-starred with John David Washington") and 2) "recalling" this information at the final token position before generating a target fact. In some cases, models need information from both of these pathways to correctly generate finetuned facts while, in other cases, either the "enrichment" or "recall" pathway alone is sufficient. We localize the "recall'' pathway to model components---finding that "recall" occurs via both task-specific attention mechanisms and an entity-specific extraction step in the feedforward networks of the final layers before the target prediction. By targeting model components and parameters, as opposed to just activations, we are able to understand the mechanisms by which finetuned knowledge is retrieved during generation.

Automated theorem proving (ATP) --- the task of generating a proof that passes automated proof verification given a math question in formal language --- is a critical challenge at the intersection of mathematics and Artificial Intelligence (AI). We introduce Goedel-Prover-V2, a family of two language models that establish a new state-of-the-art (SOTA) in open-source ATP, using the Lean proof assistant. In addition to standard expert iteration and reinforcement learning, our approach incorporates three key innovations: (1) During training when improvement plateaus on human questions, the prover does scaffolded data synthesis to generate synthetic questions of increasing difficulty for its own training; (2) The prover is trained to self-correct using Lean compiler feedback; (3) Improved test-time exploration through checkpoint averaging to balance accuracy and diversity. Our small model, Goedel-Prover-V2-8B, reaches 84.6\% pass@32 on MiniF2F and outperforms DeepSeek-Prover-V2-671B despite being $80\times$ smaller. Our flagship model, Goedel-Prover-V2-32B, achieves 88.1\% on MiniF2F at pass@32 in standard mode and 90.4\% in self-correction mode, outperforming prior SOTA by a large margin. Additionally, our flagship model solves 86 problems on PutnamBench at pass@184, securing first place among open-source models and surpassing DeepSeek-Prover-V2-671B's record of 47 problems by pass@1024 with about $20\times$ smaller model size and significantly lower compute budget. Our models, code, and data are released at \url{https://github.com/Goedel-LM/Goedel-Prover-V2}.

It is a central challenge in deep learning to understand how neural networks learn representations. A leading approach is the Neural Feature Ansatz (NFA) (Radhakrishnan et al., 2024), a conjectured mechanism for how feature learning occurs. Although the NFA is empirically validated, it is an educated guess and lacks a theoretical basis, and thus it is unclear when it might fail, and how to improve it. In this paper, we take a first-principles approach to understanding why this observation holds, and when it does not. We use first-order optimality conditions to derive the Features at Convergence Theorem (FACT), an alternative to the NFA that (a) obtains greater agreement with learned features at convergence, (b) explains why the NFA holds in most settings, and (c) captures essential feature learning phenomena in neural networks such as grokking behavior in modular arithmetic and phase transitions in learning sparse parities, similarly to the NFA. Thus, our results unify theoretical first-order optimality analyses of neural networks with the empirically-driven NFA literature, and provide a principled alternative that provably and empirically holds at convergence.

Modeling the relationships among variables has become increasingly important, particularly in high-dimensional multivariate time series forecasting tasks. However, most existing methods primarily focus on capturing coarse-grained correlations between variables, overlooking a finer and more dynamic aspect: the variable interactions often manifest differently as time progresses. To address this limitation, we propose FACT, an Fine-grained Across-variable Convolution architecture for multivariate Time series forecasting that explicitly models fine-grained variable interactions from both the time and frequency domains. Technically, we introduce a depth-wise convolution block DConvBlock, which leverages a depth-wise convolution architecture with channel-specific kernels to model dynamic variable interactions at each granularity. To further enhance efficiency, we reconfigure the original one-dimensional variables into a two-dimensional space, reducing the variable distance and the required model layers. Then DConvBlock incorporates multi-dilated 2D convolutions with progressively increasing dilation rates, enabling the model to capture fine-grained and dynamic variable interactions while efficiently attaining a global reception field. Extensive experiments on twelve benchmark datasets demonstrate that FACT not only achieves state-of-the-art forecasting accuracy but also delivers substantial efficiency gains, significantly reducing both training time and memory consumption compared to attention mechanism. The code is available at https://anonymous.4open.science/r/FACT-MTSF.

We study the problem of sampling from strongly log-concave distributions over $\mathbb{R}^d$ using the Poisson midpoint discretization (a variant of the randomized midpoint method) for overdamped/underdamped Langevin dynamics. We prove its convergence in the 2-Wasserstein distance ($\mathcal W_2$), achieving a cubic speedup in dependence on the target accuracy ($\epsilon$) over the Euler-Maruyama discretization, surpassing existing bounds for randomized midpoint methods. Notably, in the case of underdamped Langevin dynamics, we demonstrate the complexity of $\mathcal W_2$ convergence is much smaller than the complexity lower bounds for convergence in $L^2$ strong error established in the literature.

Large Language Models (LLMs) have demonstrated remarkable success across various domains, yet their optimization remains a significant challenge due to the complex and high-dimensional loss landscapes they inhabit. While adaptive optimizers such as AdamW are widely used, they suffer from critical limitations, including an inability to capture interdependencies between coordinates and high memory consumption. Subsequent research, exemplified by SOAP, attempts to better capture coordinate interdependence but incurs greater memory overhead, limiting scalability for massive LLMs. An alternative approach aims to reduce memory consumption through low-dimensional projection, but these methods lose the gradient information in the residual space, resulting in less effective optimization. In this paper, we propose COSMOS, a novel hybrid optimizer that leverages the varying importance of eigensubspaces in the gradient matrix to achieve memory efficiency without compromising optimization performance. The design of COSMOS is motivated by our empirical insights and practical considerations. Specifically, COSMOS applies SOAP to the leading eigensubspace, which captures the primary optimization dynamics, and MUON to the remaining eigensubspace, which is less critical but computationally expensive to handle with SOAP. This hybrid strategy significantly reduces memory consumption while maintaining robust optimization performance, making it particularly suitable for massive LLMs. Numerical experiments on various datasets and transformer architectures are provided to demonstrate the effectiveness of COSMOS.