ICLR 2026 论文列表

Masked diffusion models (MDMs) have recently emerged as a promising alternative to autoregressive models over discrete domains. MDMs generate sequences in an any-order, parallel fashion, enabling fast inference and strong performance on non-causal tasks. However, a crucial limitation is that they do not support token insertions and are thus limited to *fixed-length* generations. To this end, we introduce **Flex**ible **M**asked **D**iffusion **M**odels (FlexMDMs), a discrete diffusion paradigm that simultaneously can model sequences of flexible length while provably retaining MDMs' flexibility of any-order inference. Grounded in an extension of the stochastic interpolant framework, FlexMDMs generate sequences by inserting mask tokens and unmasking them. Empirically, we show that FlexMDMs match MDMs in perplexity while modeling length statistics with much higher fidelity. On a synthetic maze planning task, they achieve $\approx$ 60\% higher success rate than MDM baselines. Finally, we show pretrained MDMs can easily be *retrofitted* into FlexMDMs: on 16 H100s, it takes only three days to fine-tune LLaDA-8B into a FlexMDM, achieving superior performance on math (GSM8K, 58\%$\to$67\%) and code infilling performance (52\%$\to$65\%).

Reasoning models leveraging long chains of thought employ various cognitive skills, such as verification of their answers, backtracking, retrying by an alternate method, and more. Previous work has shown that when a base language model exhibits these skills, training that model further with reinforcement learning (RL) can learn to leverage them. How can we get models to leverage skills that aren't exhibited by base models? Our work, SkillFactory, is a method for fine-tuning models to roughly learn these skills during a supervised fine-tuning (SFT) stage prior to RL. Our approach does not rely on distillation from a stronger model, but instead uses samples from the model itself, rearranged to provide training data in the format of those skills. These "silver" SFT traces may be imperfect, but are nevertheless effective for priming a model to acquire skills during RL. Our evaluation shows that (1) starting from SkillFactory SFT initialization helps a model to generalize to harder variants of a task post-RL, despite lower performance pre-RL; (2) cognitive skills are indeed used by the model; (3) RLed SkillFactory models are more robust to regression on out-of-domain tasks than RLed base models. Our work suggests that inductive biases learned prior to RL help models learn robust cognitive skill use.

To operate effectively in the real world, robots should integrate multimodal reasoning with precise action generation. However, existing vision-language-action (VLA) models often sacrifice one for the other, narrow their abilities to task-specific manipulation data, and suffer catastrophic forgetting of pre-trained vision-language capabilities. To bridge this gap, we introduce **InstructVLA**, an end-to-end VLA model that preserves the flexible reasoning of large vision-language models (VLMs) while delivering leading manipulation performance with the help of embodied reasoning. InstructVLA introduces a novel training paradigm, *Vision-Language-Action Instruction Tuning (VLA-IT)*, which employs multimodal training with mixture-of-experts adaptation to jointly optimize embodied reasoning and action generation on both standard VLM corpora and a curated 650K-sample VLA-IT dataset. On in-domain SimplerEnv tasks, InstructVLA achieves 33.3% improvement over SpatialVLA. To evaluate generalization, we introduce SimplerEnv-Instruct, an 80-task benchmark requiring closed-loop control and high-level instruction understanding, where it outperforms a fine-tuned OpenVLA by 96% and an action expert aided by GPT-4o by 29%. Additionally, InstructVLA surpasses baseline VLMs on multimodal tasks and exhibits inference-time scaling by leveraging textual reasoning to boost manipulation performance in both simulated and real-world settings. These results demonstrate InstructVLA's potential for bridging intuitive and steerable human-robot interaction with efficient policy learning.

Reinforcement Learning (RL) has emerged as a popular training paradigm, particularly when paired with reasoning models. While effective, it primarily focuses on generating responses and lacks mechanisms to explicitly foster critique or reflection. Several recent studies, like Critique-Fine-Tuning (CFT) and Critique-Guided-Distillation (CGD) have shown the benefits of explicitly teaching LLMs how to critique. Motivated by them, we propose Critique Reinforcement Learning (CRL), where the model is tasked with generating a critique for a given (question, solution) pair. The reward is determined solely by whether the final judgment label $c \in \{\texttt{True}, \texttt{False}\}$ of the generated critique aligns with the ground-truth judgment $c^*$. Building on this point, we introduce Critique-Coder, which is trained on a hybrid of RL and CRL by substituting 20\% of the standard RL data with CRL data. We fine-tune multiple models (Critique-Coder) and evaluate them on different benchmarks to show their advantages over RL-only models. We show that Critique-Coder consistently outperforms RL-only baselines on all the evaluated benchmarks. Notably, our Critique-Coder-8B can reach over 60\% on LiveCodeBench (v5), outperforming other reasoning models like DeepCoder-14B and GPT-o1. Beyond code generation, Critique-Coder also demonstrates enhanced general reasoning abilities, as evidenced by its better performance on logic reasoning tasks from the BBEH dataset. This indicates that the application of CRL on coding datasets enhances general reasoning and critique abilities, which are transferable across a broad range of tasks. Hence, we believe that CRL works as a great complement to standard RL for LLM reasoning.

Language models serve as proxies for human preference judgements in alignment and evaluation, yet they exhibit systematic miscalibration, prioritizing superficial patterns over substantive qualities. This bias manifests as overreliance on features like length, structure, and style, leading to issues like reward hacking and unreliable evaluations. However, the connection between training data artifacts and the miscalibrated preferences exhibited by models remains poorly understood. In this work, we systematically investigate the relationship between training data biases and preference model miscalibration across five idiosyncratic features of language model generations: length, structure, jargon, sycophancy and vagueness. Using controlled counterfactual pairs, we first quantify the extent to which preference models favor responses with artificially magnified biases (\textit{skew}), finding this preference occurs in $>60$\% of instances, and model preferences show high \textit{miscalibration} ($\approx 40$\%) compared to human preferences. Notably, bias features only show mild negative correlations to human preference labels (mean $r_{\mathrm{human}} = -0.12$) but show moderately strong positive correlations with labels from a strong reward model (mean $r_{\mathrm{model}} = +0.36$), suggesting that models may overrely on spurious cues. To mitigate these issues, we propose a simple post-training method based on counterfactual data augmentation (CDA) using synthesized contrastive examples. Fine-tuning models with CDA reduces average miscalibration from 39.4\% to 32.5\% and average absolute skew difference from 20.5\% to 10.0\%, while maintaining overall RewardBench performance, indicating that targeted debiasing can strengthen the reliability of preference models within standard alignment pipelines.

We introduce FragFM, a novel hierarchical framework via fragment-level discrete flow matching for efficient molecular graph generation. FragFM generates molecules at the fragment level, leveraging a coarse-to-fine autoencoder to reconstruct details at the atom level. Together with a stochastic fragment bag strategy to effectively handle an extensive fragment space, our framework enables more efficient and scalable molecular generation. We demonstrate that our fragment-based approach achieves better property control than the atom-based method and additional flexibility through conditioning the fragment bag. We also propose a Natural Product Generation benchmark (NPGen) to evaluate modern molecular graph generative models' ability to generate natural product-like molecules. Since natural products are biologically prevalidated and differ from typical drug-like molecules, our benchmark provides a more challenging yet meaningful evaluation relevant to drug discovery. We conduct a FragFM comparative study against various models on diverse molecular generation benchmarks, including NPGen, demonstrating superior performance. The results highlight the potential of fragment-based generative modeling for large-scale, property-aware molecular design, paving the way for more efficient exploration of chemical space.

Large pre-trained models achieve remarkable success across diverse domains, yet fully fine-tuning incurs prohibitive computational and memory costs. Parameter-efficient fine-tuning (PEFT) has thus become a mainstream paradigm. Among them, Low-Rank Adaptation (LoRA) introduces trainable low-rank matrices and shows strong performance, nevertheless, its fixed-rank design limits flexibility. Dynamic rank allocation methods mitigate this issue by pruning redundant directions; however, they often rely on heuristic, element-level metrics that globally sort rank directions without matrix-wise distinction, and they lack mechanisms to expand capacity in layers requiring additional adaptation. To overcome these limitations, we propose FlexLoRA, an entropy-guided flexible low-rank adaptation framework that (i) evaluates matrix importance via spectral energy entropy, (ii) supports rank pruning and expansion under a global budget, and (iii) employs zero-impact initialization for newly added singular directions to ensure stability. By addressing granularity, flexibility, and stability limitations, FlexLoRA provides a more principled solution for PEFT. Extensive experiments show that FlexLoRA consistently outperforms state-of-the-art baselines across benchmarks.

Neural radiance fields and 3D Gaussian splatting (3DGS) have significantly advanced 3D reconstruction and novel view synthesis (NVS). Yet, achieving high-fidelity and view-consistent renderings directly from point clouds---without costly per-scene optimization---remains a core challenge. In this work, we present DiffPBR, a diffusion-based framework that synthesizes coherent, photorealistic renderings from diverse point cloud inputs. We demonstrate that diffusion models, when guided by viewpoint-projected noise explicitly constrained by scene geometry and visibility, naturally enforce geometric consistency across camera motion. To achieve this, we first introduce adaptive CoNo-Splatting, a technique for fast and faithful rasterization that ensures efficient and effective handling of point clouds. Secondly, we integrate residual learning into the neural re-rendering pipeline, which improves convergence, generalization, and visual quality across diverse rendering tasks. Extensive experiments show that our method outperforms existing baselines with an improvement of **3~5dB** in rendered image quality, a reduction from **41 to 8** in GPU hours for training, and an increase from **3.6fps to 10fps** (our one-step variant) in rendering speed frequency.

Unsupervised representation learning, particularly sequential disentanglement, aims to separate static and dynamic factors of variation in data without relying on labels. This remains a challenging problem, as existing approaches based on variational autoencoders and generative adversarial networks often rely on multiple loss terms, complicating the optimization process. Furthermore, sequential disentanglement methods face challenges when applied to real-world data, and there is currently no established evaluation protocol for assessing their performance in such settings. Recently, diffusion models have emerged as state-of-the-art generative models, but no theoretical formalization exists for their application to sequential disentanglement. In this work, we introduce the Diffusion Sequential Disentanglement Autoencoder (DiffSDA), a novel, modal-agnostic framework effective across diverse real-world data modalities, including time series, video, and audio. DiffSDA leverages a new probabilistic modeling, latent diffusion, and efficient samplers, while incorporating a challenging evaluation protocol for rigorous testing. Our experiments on diverse real-world benchmarks demonstrate that DiffSDA outperforms recent state-of-the-art methods in sequential disentanglement.

Real-world physical systems are inherently complex, often involving the coupling of multiple physics, making their simulation both highly valuable and challenging. Many mainstream approaches face challenges when dealing with decoupled data. Besides, they also suffer from low efficiency and fidelity in strongly coupled spatio-temporal physical systems. Here we propose GenCP, a novel and elegant generative paradigm for coupled multiphysics simulation. By formulating coupled-physics modeling as a probability modeling problem, our key innovation is to integrate probability density evolution in generative modeling with iterative multiphysics coupling, thereby enabling training on data from decoupled simulation and inferring coupled physics during sampling. We also utilize operator-splitting theory in the space of probability evolution to establish error controllability guarantees for this “conditional-to-joint” sampling scheme. We evaluate our paradigm on a synthetic setting and three challenging multiphysics scenarios to demonstrate both principled insight and superior application performance of GenCP. Code is available at this repo: https://github.com/AI4Science-WestlakeU/GenCP.

Brain tumor segmentation in multi-modal MRIs poses significant challenges when one or more modalities are missing. Recent approaches commonly employ parallel fusion strategies; however, these methods often risk losing crucial shared information across modalities, which can degrade segmentation performance. In this paper, we advocate leveraging sequential information bottleneck fusion to effectively preserve shared information across modalities. From an information-theoretic perspective, sequential fusion not only produces more robust fused representations in missing-data scenarios but also achieves a tighter generalization upper bound compared to parallel fusion approaches. Building on this principle, we propose the Sequential Multi-modal Segmentation Network (SMSN), which integrates an Information-Bottleneck Fusion Module (IBFM). The IBFM sequentially extracts modality-common features while reconstructing modality-specific features through a dedicated feature extraction module. Extensive experiments on the BRATS18 and BRATS20 glioma datasets demonstrate that SMSN consistently outperforms traditional parallel fusion-based baselines, achieving exceptional robustness in diverse missing-modality settings. Furthermore, SMSN exhibits superior cross-domain generalization, as evidenced by its ability to transfer a trained model from BRATS20 to a brain metastasis dataset without fine-tuning. To ensure reproducibility, the code of the SMSN is provided in the supplementary file.

The rapid progress of visual autoregressive (VAR) models has brought new opportunities for text-to-image generation, but also heightened safety concerns. Existing concept erasure techniques, primarily designed for diffusion models, fail to generalize to VARs due to their next-scale token prediction paradigm. In this paper, we first propose a novel VAR Erasure framework **VARE** that enables stable concept erasure in VAR models by leveraging auxiliary visual tokens to reduce fine-tuning intensity. Building upon this, we introduce **S-VARE**, a novel and effective concept erasure method designed for VAR, which incorporates a filtered cross entropy loss to precisely identify and minimally adjust unsafe visual tokens, along with a preservation loss to maintain semantic fidelity, addressing the issues such as language drift and reduced diversity introduce by na\"ive fine-tuning. Extensive experiments demonstrate that our approach achieves surgical concept erasure while preserving generation quality, thereby closing the safety gap in autoregressive text-to-image generation by earlier methods.

Recent advances in multimodal large language models (MLLMs) have opened new opportunities for embodied intelligence, enabling multimodal understanding, reasoning, and interaction, as well as continuous spatial decision-making. Nevertheless, current MLLM-based embodied systems face two critical limitations. First, **Geometric Adaptability Gap:** models trained solely on 2D inputs or with hard-coded 3D geometry injection suffer from either insufficient spatial information or restricted 2D generalization, leading to poor adaptability across tasks with diverse spatial demands. Second, **Embodiment Constraint Gap**: prior work often neglects the physical constraints of real robots, resulting in task plans that are theoretically valid but practically infeasible.To address these gaps, we introduce **OmniEVA** -- an embodied versatile planner that enables advanced embodied reasoning and task planning through two pivotal innovations: (1) a **Task-Adaptive 3D Grounding** mechanism, which uses a gated router to dynamically inject 3D features based on task context, enabling selective geometric reasoning. (2) an **Embodiment-Aware Reasoning** framework that incorporates task goals and physical constraints into the reasoning loop, ensuring executable plans. Extensive experiments show that OmniEVA achieves state-of-the-art performance on 7 of 8 embodied reasoning benchmarks and excels in downstream tasks such as object navigation and mobile manipulation. Evaluations on proposed primitive and composite benchmarks confirm its robust and versatile planning capabilities.

Vision-Language models (VLMs), including CLIP, are known to encode biases such as learning spurious correlations that falsely associate background attributes with particular labels. Debiasing approaches typically aim to isolate and remove subspaces corresponding to a target concept via projecting its embedding away from the concept. This strategy succeeds in debiasing VLM embeddings with respect to the concepts considered but can amplify biased shortcuts in unconsidered concepts. In practice, it is impossible to enumerate all possible biases, meaning that an increase in bias can go unobserved during evaluation. We propose a debiasing approach for a set of known concepts such that the relation to the remaining, unconsidered, concepts is minimally changed. We achieve this by rotating the VLM’s embeddings within only a relevant subspace, rather than removing these subspaces, which mitigates unintended bias amplification.

We introduce VL-JEPA, a vision-language model built on a Joint Embedding Predictive Architecture (JEPA). Instead of autoregressively generating tokens as in classical VLMs, VL-JEPA predicts continuous embeddings of the target texts. By learning in an abstract representation space, the model can focus on task-relevant semantics while abstracting away surface-level linguistic variability. In a strictly controlled comparison against standard token-space VLM training with the same vision encoder and training data, VL-JEPA achieves stronger performance while having 50% fewer trainable parameters. At inference time, a lightweight text decoder is invoked only when needed to translate VL-JEPA predicted embeddings into text. We show that VL-JEPA natively supports selective decoding that can reduce the number of decoding operations by approximately 2.85× while maintaining similar performance compared to dense non-adaptive uniform decoding. Beyond generation, the embedding-space formulation naturally supports open-vocabulary classification, text-to-video retrieval, and discriminative VQA without any architecture modification. On eight video classification and eight video retrieval datasets, the average performance of VL-JEPA surpasses that of CLIP, SigLIP2, and Perception Encoder. At the same time, the model achieves comparable performance to classical VLMs (InstructBLIP, QwenVL) on four VQA datasets: GQA, TallyQA, POPE, and POPEv2, despite only having 1.6B parameters.

Query-based image retrieval (QBIR) requires retrieving relevant images given diverse and often stylistically heterogeneous queries, such as sketches, artworks, or low-resolution previews. While large-scale vision--language representation models (VLRMs) like CLIP offer strong zero-shot retrieval performance, they struggle with distribution shifts caused by unseen query styles. In this paper, we propose the Hypernetwork-driven Style-adaptive Retrieval (Hystar), a lightweight framework that dynamically adapts model weights to each query’s style. Hystar employs a hypernetwork to generate singular-value perturbations ($\Delta S$) for attention layers, enabling flexible per-input adaptation, while static singular-value offsets on MLP layers ensure cross-style stability. To better handle semantic confusions across styles, we design StyleNCE as part of Hystar, an optimal-transport-weighted contrastive loss that emphasizes hard cross-style negatives. Extensive experiments on multi-style retrieval and cross-style classification benchmarks demonstrate that Hystar consistently outperforms strong baselines, achieving state-of-the-art performance while being parameter-efficient and stable across styles.

Federated Learning (FL) enables decentralized model training across clients without sharing raw data, but its performance degrades under real-world data heterogeneity. Existing methods often fail to address distribution shift across clients and distribution drift over time, or they rely on unrealistic assumptions such as known number of client clusters and data heterogeneity types, which limits their generalizability. We introduce **Feroma**, a novel FL framework that explicitly handles both distribution shift and drift without relying on client or cluster identity. **Feroma** builds on client distribution profiles—compact, privacy-preserving representations of local data—that guide model aggregation and test-time model assignment through adaptive similarity-based weighting. This design allows **Feroma** to dynamically select aggregation strategies during training, ranging from clustered to personalized, and deploy suitable models to unseen, and unlabeled test clients without retraining, online adaptation, or prior knowledge on clients' data. Extensive experiments show that compared to 10 state-of-the-art methods, **Feroma** improves performance and stability under dynamic data heterogeneity conditions—an average accuracy gain of up to 12 percentage points over the best baselines across 6 benchmarks—while maintaining computational and communication overhead comparable to FedAvg. These results highlight that distribution-profile-based aggregation offers a practical path toward robust FL under both data distribution shifts and drifts.

Multinomial logistic regression (MLR) is a classic model for multi-class classification that has been widely used for neural decoding. However, MLR requires a finite set of discrete output classes, limiting its applicability to settings with continuous-valued outputs (e.g., time, orientation, velocity, or spatial position). To address this limitation, we propose Continuous Multinomial Logistic Regression (CMLR), a generalization of MLR to continuous output spaces. CMLR defines a novel exponential-family model for conditional density estimation (CDE), mapping neural population activity to a full probability density over external covariates. It captures the influence of each neuron’s activity on the decoded variable through a smooth, interpretable tuning function, regularized by a Gaussian process prior. The resulting nonparametric decoding model flexibly captures a wide variety of conditional densities, including multimodal, asymmetric, and circular distributions. We apply CMLR to large-scale datasets from mouse and monkey primary visual cortex, mouse hippocampus, and monkey motor cortex, and show that it consistently outperforms a wide variety of other decoding methods, including deep neural networks (DNNs), XGBoost, and FlexCode. It also outperforms correlation-blind models such as Naive Bayes, highlighting the importance of correlations for accurate neural decoding. The CMLR model provides a scalable, flexible, and interpretable method for decoding responses from diverse brain regions.

Current methods for multivariate time series forecasting can be classified into channel-dependent and channel-independent models. Channel-dependent models learn cross-channel features but often overfit the channel ordering, which hampers adaptation when channels are added or reordered. Channel-independent models treat each channel in isolation to increase flexibility, yet this neglects inter-channel dependencies and limits performance. To address these limitations, we propose CPiRi, a channel permutation invariant (CPI) framework that infers cross-channel structure from data rather than memorizing a fixed ordering, enabling deployment in settings with structural and distributional co-drift without retraining. CPiRi couples spatio-temporal decoupling architecture with permutation-invariant regularization training strategy: a frozen pretrained temporal encoder extracts high-quality temporal features, a lightweight spatial module learns content-driven inter-channel relations, while a channel shuffling strategy enforces CPI during training. We further ground CPiRi in theory by analyzing permutation equivariance in multivariate time series forecasting. Experiments on multiple benchmarks show state-of-the-art results. CPiRi remains stable when channel orders are shuffled and exhibits strong inductive generalization to unseen channels even when trained on only half of the channels, while maintaining practical efficiency on large-scale datasets. The source code is released at https://github.com/JasonStraka/CPiRi.