ICLR 2026 论文列表

Large reasoning models (LRMs), such as OpenAI's o1 and DeepSeek-R1, harness test-time scaling to perform multi-step reasoning for complex problem-solving. This reasoning process, executed before producing final answers, is often guided by special juncture tokens that prompt self-evaluative reflection. We refer to these transition markers and reflective cues as “reflection tokens” (e.g., “wait,” “but,” “alternatively”). In this work, we treat reflection tokens as a “resource” and introduce the problem of resource allocation, aimed at improving the test-time compute performance of LRMs by adaptively regulating the frequency and placement of reflection tokens. Through empirical analysis, we show that both excessive and insufficient use of reflection tokens, referred to as over-reflection and under-reflection, can degrade model performance. To better understand this trade-off, we draw an analogy between reflection token usage and learning rate scheduling in optimization. Building on this insight, we propose cyclical reflection token scheduling (termed Ours), a training-free decoding strategy that dynamically modulates reflection token logits with a bidirectional, position-dependent triangular waveform, incurring no additional computation cost. Experiments on MATH500, AIME2024/2025, AMC2023, GPQA Diamond, and LiveCodeBench demonstrate that Ours consistently improves performance across model sizes (1.5B–8B), outperforming standard decoding and recent approaches such as TIP (thought switching penalty) and S1.

Many modern biological assays, including RNA sequencing, yield integer-valued counts that reflect the number of molecules detected. These measurements are often not at the desired resolution: while the unit of interest is typically a single cell, many measurement technologies produce counts aggregated over sets of cells. Although recent generative frameworks such as diffusion and flow matching have been extended to non-Euclidean and discrete settings, it remains unclear how best to model integer-valued data or how to systematically deconvolve aggregated observations. We introduce Count Bridges, a stochastic bridge process on the integers that provides an exact, tractable analogue of diffusion-style models for count data, with closed-form conditionals for efficient training and sampling. We extend this framework to enable direct training from aggregated measurements via an Expectation-Maximization approach that treats unit-level counts as latent variables. We demonstrate state-of-the-art performance on integer distribution matching benchmarks, comparing against flow matching and discrete flow matching baselines across various metrics. We then apply Count Bridges to two large-scale problems in biology: modeling single-cell gene expression data at the nucleotide resolution, with applications to deconvolving bulk RNA-seq, and resolving multicellular spatial transcriptomic spots into single-cell count profiles. Our methods offer a principled foundation for generative modeling and deconvolution of biological count data across scales and modalities.

Reinforcement Learning with Verifiable Rewards (RLVR), particularly with algorithms like Group Relative Policy Optimization (GRPO), has proven highly effective in enhancing the reasoning capabilities of large language models. However, a critical bottleneck in current pipelines lies in the limited diversity of sampled trajectories during group rollouts. Homogeneous trajectories and their associated rewards would diminish the return signals for policy updates, thereby hindering effective policy learning. This lack of diversity stems primarily from token-level stochastic sampling, where local variations are likely to collapse into near-identical reasoning paths. To address this limitation, we propose Lookahead Tree-Based Rollouts (LATR), a novel rollout strategy designed to explicitly promotes trajectory-level diversity by enforcing branching into different candidate tokens likely to yield distinct continuations. Specifically, LATR iteratively operates in three stages: (1) branching at high-uncertainty generation steps, (2) performing lookahead simulation for each new branch, and (3) pruning branches that exhibits prolonged similarity during simulation. Compared with Stochastic Sampling, LATR accelerates policy learning by 131% on average and improves final pass@1 performance by 4.2% on both GRPO and Dynamic sAmpling Policy Optimization (DAPO) algorithms across different reasoning tasks. Our code will be publicly available.

Direct Preference Optimization (DPO) has emerged as a popular algorithm for aligning pretrained large language models with human preferences, owing to its simplicity and training stability. However, DPO suffers from the recently identified squeezing effect (also known as likelihood displacement), where the probability of preferred responses decreases unintentionally during training. To understand and mitigate this phenomenon, we develop a theoretical framework that models the coordinate-wise dynamics in the logit space. Our analysis reveals that negative-gradient updates cause residuals to expand rapidly along high-curvature directions, which underlies the squeezing effect, whereas Sharpness-Aware Minimization (SAM) can suppress this behavior through its curvature-regularization effect. Building on this insight, we investigate logits-SAM, a computationally efficient variant that perturbs only the output layer with negligible overhead. Extensive experiments on Pythia-2.8B and Mistral-7B across multiple datasets demonstrate that logits-SAM consistently improves the effectiveness of DPO.

We propose LayerSync, a domain-agnostic approach for improving the generation quality and the training efficiency of diffusion models. Prior studies have highlighted the connection between the quality of generation and the representations learned by diffusion models, showing that external guidance on model intermediate representations accelerates training. We reconceptualize this paradigm by regularizing diffusion models with their own intermediate representations. Building on the observation that representation quality varies across diffusion model layers, we show that the most semantically rich representations can act as an intrinsic guidance for weaker ones, reducing the need for external supervision. Our approach, LayerSync, is a self-sufficient, plug-and-play regularizer term with no overhead on diffusion model training and generalizes beyond the visual domain to other modalities. LayerSync requires no pretrained models nor additional data. We extensively evaluate the method on image generation and demonstrate its applicability to other domains such as audio, video, and motion generation. We show that it consistently improves the generation quality and the training efficiency. For example, we speed up the training of flow-based transformer by over 8.75$\times$ on ImageNet dataset and improve the generation quality by 23.6\%.

The transformer architecture has demonstrated remarkable capabilities in modern artificial intelligence, among which the capability of implicitly learning an internal model during inference time is widely believed to play a key role in the understanding of pre-trained large language models. However, most recent works have been focusing on studying supervised learning topics such as in-context learning, leaving the field of unsupervised learning largely unexplored. This paper investigates the capabilities of transformers in solving Gaussian Mixture Models (GMMs), a fundamental unsupervised learning problem through the lens of statistical estimation. We propose a transformer-based learning framework called Transformer for Gaussian Mixture Models (TGMM) that simultaneously learns to solve multiple GMM tasks using a shared transformer backbone. The learned models are empirically demonstrated to effectively mitigate the limitations of classical methods such as Expectation-Maximization (EM) or spectral algorithms, at the same time exhibit reasonable robustness to distribution shifts. Theoretically, we prove that transformers can efficiently approximate both the Expectation-Maximization (EM) algorithm and a core component of spectral methods—namely, cubic tensor power iterations. These results not only improve upon prior work on approximating the EM algorithm, but also provide, to our knowledge, the first theoretical guarantee that transformers can approximate high-order tensor operations. Our study bridges the gap between practical success and theoretical understanding, positioning transformers as versatile tools for unsupervised learning.

Recent advances in Open Vocabulary Object Detection (OVOD) have shown strong performance on standard benchmarks, but performance drops sharply under out-of-distribution (OOD) shifts. Continual learning offers a potential remedy by sequentially integrating new tasks, yet existing methods often struggle to balance retaining the pre-trained model capabilities with adapting to new tasks, and usually require retraining under specific task orders. To address these limitations, we observe that model editing naturally lends itself to this setting, as it enables efficient knowledge injection while retaining prior capabilities. Building on this insight, we introduce $\textbf{A}$utomatically $\textbf{B}$alanced $\textbf{M}$odel $\textbf{E}$diting ($\textbf{ABME}$), which injects new task knowledge into the powerful OVOD models while preserving the model’s original abilities. We first stores compact key–value representations with storage cost independent of task volume. Then we leverage the stored KV matrices to automatically balance the new and old knowledge for varying learning scenarios, supporting order-agnostic task insertion or removal without additional retraining. Experiments show that ABME consistently achieves a better trade-off between maintaining pre-trained performance and adapting to diverse OOD tasks compared to existing continual learning approaches for open-vocabulary object detection, and generalizes seamlessly across different models and task scales.

The common approach for developing a vision model is generalism, which involves training on a large diverse dataset to cover the varied deployment environments and leads to a model that is expected to solve the problem everywhere. However, many practical applications need to operate in a specific test space, e.g., a robot deployed in a single house, and do not necessarily need to generalize to novel environments. In this work, we explore whether we can use rich multimodal data only from the test environment to pre-train a representation in a self-supervised way, without access to any external data. We find that this approach can match and, in most cases, outperform generalists pre-trained on large-scale Internet datasets, including popular off-the-shelf models, CLIP and DINOv2. We study the effectiveness of this approach by evaluating the models on various datasets and downstream tasks, such as semantic segmentation, captioning, and object detection, as well as a set of ablations and analyses to extract insights. This approach raises intriguing points on substituting data with (multi)modality, enabling an alternative scenario where the need for external Internet-scale datasets for pre-training models is reduced. It also shows that merely benefiting from test-space data was insufficient for achieving competitive results, and multimodality was essential for that purpose.

In recent years, Contrastive Learning (CL) has garnered significant attention due to its efficacy across various domains, spanning from visual and textual modalities. A fundamental aspect of CL is aligning the representations of anchor instances with relevant positive samples while simultaneously separating them from negative ones. Prior studies have extensively explored diverse strategies for generating and sampling contrastive (i.e., positive/negative) pairs. Despite the empirical success, the theoretical understanding of the CL approach remains under-explored, leaving questions such as the rationale behind contrastive-pair sampling and its contributions to the model performance unclear. This paper addresses this gap by providing a comprehensive theoretical analysis from the angle of optimality conditions and introducing the SnaPhArd Contrast Learning (SPACL). Specifically, SPACL prioritizes hard positive and hard negative samples during constructing contrastive pairs and computing the contrastive loss, rather than treating all samples equally. Experimental results across two downstream tasks demonstrate that SPACL consistently outperforms or competes favorably with state-of-the-art methods, showcasing its robustness and efficacy. A comprehensive ablation study further examines the effectiveness of SPACL's individual components to verify the theoretic findings.

Unified multimodal models integrate the reasoning capacity of large language models with both image understanding and generation, showing great promise for advanced multimodal intelligence. However, the community still lacks a rigorous reasoning-centric benchmark to systematically evaluate the alignment between understanding and generation, and their generalization potential in complex visual tasks. To this end, we introduce \textbf{GIR-Bench}, a comprehensive benchmark that evaluates unified models across three complementary perspectives. Firstly, we explore whether models can consistently leverage the same knowledge for both understanding and generation (GIR-Bench-Uni). Secondly, we investigate whether models can perform reasoning-centric text-to-image generation that requires applying logical constraints and implicit knowledge to generate faithful visual content (GIR-Bench-T2I). Thirdly, we evaluate whether models can handle multi-step reasoning in editing (GIR-Bench-Edit). For each subset, we carefully design different task-specific evaluation pipelines tailored for each task. This enables fine-grained and interpretable evaluation while mitigating biases from the prevalent MLLM-as-a-Judge paradigm. Extensive ablations over various unified models and generation-only systems have shown that: Although unified models are more capable of reasoning-driven visual tasks, they still exhibit a persistent gap between understanding and generation. The data and code for GIR-Bench are available at \url{https://anonymous.4open.science/r/GIR-Bench-7E40}.

Existing causal discovery methods are fundamentally limited by the assumption of a static causal graph, a constraint that fails in real-world systems where causal relationships dynamically vary with underlying system parameters. This discrepancy prevents the application of causal discovery in critical domains such as industrial process control, where understanding how causal effects change is essential. We address this gap by proposing a new paradigm that moves beyond static graphs to learn functional causal representations. We introduce a framework that models each causal link not as a static weight but as a function of measurable system parameters. By representing these functions using Polynomial Chaos Expansions (PCE), we develop a tractable method to learn the complete parametric causal structure from observational data. We provide theoretical proofs for the identifiability of these functional models and introduce a novel, provably convergent learning algorithm. On a large-scale chemical reactor dataset, our method learns the dynamic causal structure with a 90.9% F1-score, nearly doubling the performance of state-of-the-art baselines and providing an interpretable model of how causal mechanisms evolve.

Spoken Question-Answering (SQA) is a core capability for useful and interactive artificial intelligence systems. Recently, several speech-language models (SpeechLMs) have been released with a specific focus on improving their SQA performance. However, a lack of controlled ablations of pretraining data processing and curation makes it challenging to understand what factors account for performance, despite substantial gains from similar studies in other data modalities. In this work, we address this gap by conducting a data-centric exploration for pretraining SpeechLMs. We focus on three questions fundamental to speech-language pretraining data: (1) how to process raw web-crawled audio content for speech-text pretraining, (2) how to construct synthetic datasets to augment web-crawled data and (3) how to interleave (text, audio) segments into training sequences. We apply the insights from our controlled data-centric ablations to pretrain a 3.8B-parameter SpeechLM, called SpeLangy, that outperforms models that are up to 3x larger by 10.2% absolute performance. We hope our findings highlight the impact of effective data curation and guide future data-centric exploration in SpeechLMs.

The rapid advancement of video diffusion models has been hindered by fundamental limitations in temporal modeling, particularly the rigid synchronization of frame evolution imposed by conventional scalar timestep variables. While task-specific adaptations and autoregressive models have sought to address these challenges, they remain constrained by computational inefficiency, catastrophic forgetting, or narrow applicability. In this work, we present \textbf{Pusa} V1.0, a versatile model that leverages \textbf{vectorized timestep adaptation (VTA)} to enable fine-grained temporal control within a unified video diffusion framework. Note that VTA is a non-destructive adaptation, which means that it fully preserves the capabilities of the base model. \textbf{Unlike conventional methods like Wan-I2V, which finetune a base text-to-video (T2V) model with abundant resources to do image-to-video (I2V), we achieve comparable results in a zero-shot manner after an ultra-efficient finetuning process based on VTA. Moreover, this method also unlocks many other zero-shot capabilities simultaneously, such as start-end frames and video extension ---all without task-specific training. Meanwhile, it keeps the T2V capability from the base model.} Mechanistic analyses also reveal that our approach preserves the foundation model's generative priors while surgically injecting temporal dynamics, avoiding the combinatorial explosion inherent to the vectorized timestep. This work establishes a scalable, efficient, and versatile paradigm for next-generation video synthesis, democratizing high-fidelity video generation for research and industry alike.

Modeling semantic and structural information from tabular data remains a core challenge for effective table understanding. Existing Table-as-Text approaches flatten tables for large language models (LLMs), but lose crucial structural cues, while Table-as-Image methods preserve structure yet struggle with precise semantics. Recent Table-as-Multimodality strategies attempt to combine textual and visual views, but they (1) statically process both modalities for every query-table pair within large multimodal LLMs (MLLMs), inevitably introducing redundancy and even conflicts, and (2) depend on costly fine-tuning of MLLMs. In light of this, we propose TableDART, a training-efficient framework that integrates multimodal views by reusing pretrained single-modality models. TableDART introduces a lightweight 2.59M-parameter MLP gating network that dynamically selects the optimal path (Text-only, Image-only, or Fusion) for each table–query pair, reducing redundancy and avoiding conflicts that arise when textual and visual views of the same table provide inconsistent cues. By routing to the most appropriate view, our framework improves both accuracy and efficiency. In addition, we propose a novel agent to mediate cross-modal knowledge integration by analyzing outputs from text- and image-based models, either selecting the best result or synthesizing a new answer through reasoning. This design avoids the prohibitive costs of full MLLM fine-tuning. Extensive experiments on seven benchmarks show that TableDART establishes new state-of-the-art performance among open-source models, surpassing the strongest baseline by an average of 4.02%. The code is available at: https://github.com/xiaobo-xing/TableDART.

Large language models (LLMs) are typically deployed under diverse memory and compute constraints. Existing approaches build model families by training each size independently, which is prohibitively expensive and provides only coarse-grained size options. In this work, we identify a novel phenomenon that we call boomerang distillation: starting from a large base model (the teacher), one first distills down to a small student and then progressively reconstructs intermediate-sized models by re-incorporating blocks of teacher layers into the student without any additional training. This process produces zero-shot interpolated models of many intermediate sizes whose performance scales smoothly between the student and teacher, often matching or surpassing pretrained or distilled models of the same size. We further analyze when this type of interpolation succeeds, showing that alignment between teacher and student through pruning and distillation is essential. Boomerang distillation thus provides a simple and efficient way to generate fine-grained model families, dramatically reducing training cost while enabling flexible adaptation across deployment environments.

ICU (Intensive Care Unit) records comprise heterogeneous multivariate time series sampled at irregular intervals with pervasive missingness, yet clinical applications demand predictive models that are both accurate and interpretable. We present our Graph Attention-based Relational Learning for Intensive Care (GARLIC) model, a novel neural network architecture that imputes missing data through a learnable exponential-decay encoder, captures inter-sensor dependencies through time-lagged summary graphs, and fuses global patterns with cross-dimensional sequential attention. All attention weights and graph edges are learned end-to-end to serve as built-in observation-, signal-, and edge-level explanations. To reconcile auxiliary reconstruction and primary classification objectives, we develop an alternating decoupled optimization scheme that stabilizes training. On three ICU benchmarks (PhysioNet 2012 \& 2019, MIMIC-III) for outcome prediction, GARLIC sets the new state of the art, significantly improving AUROC and AUPRC over best-performing baselines at comparable computational cost. Ablation studies confirm each module’s contribution, and feature-removal trials validate importance attribution fidelity through a monotonic performance drop (full > top 50\% > random 50\% > bottom 50\%). Additionally, we also demonstrate GARLIC's superiority in data imputation and classification on various datasets beyond the ICU domain, showing its generalizability and applicability to broader tasks and datasets. Finally, real-time case studies demonstrate actionable risk warnings with transparent explanations, marking a significant advancement toward accurate, explainable deep learning for irregularly sampled ICU time series data.

Accurate simulation of flow fields around tandem geometries is critical for engineering design but remains computationally intensive. Existing machine learning approaches typically focus on simpler cases and lack evaluation on multi-body configurations. To support research in this area, we present **TandemFoilSet**: five tandem-airfoil datasets (4152 tandem-airfoil simulations) paired with four single-airfoil counterparts, for a total of 8104 CFD simulations. We provide benchmark results of a curriculum learning framework using a directional integrated distance representation, residual pre-training, training schemes based on freestream conditions and smooth-combined estimated fields, and a domain decomposition strategy. Evaluations demonstrate notable gains in prediction accuracy. We believe these datasets will enable future work on scalable, data-driven flow prediction for tandem-airfoil scenarios.

Low-Rank Adaptation (LoRA) benefits from its plug-and-play nature, enabling large language models (LLMs) to achieve significant performance gains at low cost, has driven the development of LoRA-sharing platforms. However, the jailbreak and backdoor concerns associated with LoRA-sharing platforms remain underexplored. Existing LoRA-based attacks primarily focus on achieving high attack success rates, while neglecting the core reason why LoRA is adopted by user, i.e. to gain downstream task capabilities. However, achieving effective attacks while preserving strong multi-task performance remains challenging, as the largely unrelated objectives tend to interfere with each other during optimization. In this paper, we propose JailbreakLoRA, a multi-task jailbreak LoRA training method that balances task utility and attack capability, it resolves training interference by uncertainty-weighting losses and mitigating gradient conflicts. Additionally, JailbreakLoRA is designed to generate an affirmative prefix upon trigger activation, exploiting inference-time hallucinations to enhance the effectiveness of jailbreak. Experimental results demonstrate that our method outperforms SOTA LoRA-based attacks, achieving a 16.0\% improvement in attack success rate while also enhancing performance on multi-downstream tasks by 16.5\% in average.

Learning to Defer (L2D) improves AI reliability in decision-critical environments, such as healthcare, by training a model to either make its own prediction or delerejector the decision to a human expert. A key challenge is adapting to unseen experts: those who were not involved during the system's training process. Current methods for this task, however, can falter when unseen experts are out-of-distribution (OOD) relative to the training population. We identify a core architectural flaw as the cause: they learn identity-conditioned policies by processing class-indexed signals in fixed coordinates, creating shortcuts that violate the problem's inherent permutation symmetry. We introduce Identity-Free Deferral (IFD), an architecture that enforces this symmetry by construction. From a few-shot context, IFD builds a query-independent Bayesian competence profile for each expert. It then supplies the deferral rejector with a low-dimensional, role-indexed state containing only structural information, such as the model's confidence in its top-ranked class and the expert's estimated skill for that same role, which obscures absolute class identities. We train IFD using an uncertainty-aware, context-only objective that removes the need for expensive query-time expert labels. We formally prove the permutation invariance of our approach, contrasting it with the generic non-invariance of standard population encoders. Experiments on medical imaging benchmarks and ImageNet-16H with real human annotators show that IFD consistently improves generalization to unseen experts, with significant gains in OOD settings, all while using fewer annotations than competing methods.