ICLR 2026 论文列表

When language models (LMs) are trained via reinforcement learning (RL) to generate natural language “reasoning chains”, their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or “hallucinate”) in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score—a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any analogous reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations—outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models.

Human cognition naturally engages with abstract and fluid concepts, whereas existing reasoning models often rely on generating discrete tokens, potentially constraining their expressive capabilities. Recent advancements aim to address this limitation by enabling large language models (LLMs) to generate soft, abstract tokens, thus facilitating reasoning within a continuous concept space. In this paper, we investigate the $\textit{Soft Thinking}$ capabilities of various LLMs through a systematic analysis of their internal behavior using a suite of probing techniques. Contrary to the prevailing belief that Soft Thinking supports parallel exploration of diverse reasoning paths, our findings reveal that $\textbf{LLMs behave as single-threaded reasoners}$—they predominantly rely on the token with the highest probability in the soft input to predict the next step. This behavior induces a greedy feedback loop that suppresses alternative reasoning paths and undermines the benefits of transmitting richer information via Soft Tokens. To address this $\textit{Greedy Pitfall}$, we propose $\textbf{Stochastic Soft Thinking}$, which introduces stochasticity to break free from the greedy tendency. Our experiments demonstrate that incorporating $\textit{randomness}$—particularly with the $\textbf{Gumbel-Softmax trick}$—can alleviate the limitations of vanilla approaches and unleash the potential of Soft Thinking, resulting in superior performance across eight reasoning benchmarks.

Computed Tomography (CT) plays a vital role in inspecting the internal structures of industrial objects. Furthermore, achieving high-quality CT reconstruction from sparse views is essential for reducing production costs. While classic implicit neural networks have shown promising results for sparse reconstruction, they are unable to leverage shape priors of objects. Motivated by the observation that numerous industrial objects exhibit rectangular structures, we propose a novel \textbf{N}eural \textbf{A}daptive \textbf{B}inning (\textbf{NAB}) method that effectively integrates rectangular priors into the reconstruction process. Specifically, our approach first maps coordinate space into a binned vector space. This mapping relies on an innovative binning mechanism based on differences between shifted hyperbolic tangent functions, with our extension enabling rotations around the input-plane normal vector. The resulting representations are then processed by a neural network to predict CT attenuation coefficients. This design enables end-to-end optimization of the encoding parameters---including position, size, steepness, and rotation---via gradient flow from the projection data, thus enhancing reconstruction accuracy. By adjusting the smoothness of the binning function, NAB can generalize to objects with more complex geometries. This research provides a new perspective on integrating shape priors into neural network-based reconstruction. Extensive experiments demonstrate that NAB achieves superior performance on two datasets.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language tasks yet remain limited in long video understanding due to the limited context window. Consequently, prevailing approaches tend to rely on uniform frame sampling or static pre-selection, which might overlook critical evidence and unable to correct its initial selection error during its reasoning process. To overcome these limitations, we propose VideoZoomer, a novel agentic framework that enables MLLMs to dynamically control their visual focus during reasoning. Starting from a coarse low-frame-rate overview, VideoZoomer invokes a temporal zoom tool to obtain high-frame-rate clips at autonomously chosen moments, thereby progressively gathering fine-grained evidence in a multi-turn interactive manner. Accordingly, we adopt a two-stage training strategy: a cold-start supervised fine-tuning phase on a curated dataset of distilled exemplar and reflection trajectories, followed by reinforcement learning to further refine the agentic policy. Extensive experiments demonstrate that our 7B model demonstrates diverse and complex reasoning patterns, yielding strong results across a broad set of long video understanding and reasoning benchmarks. These emergent capabilities allow it to consistently surpass existing open-source models and even rival proprietary systems on challenging tasks, while achieving superior efficiency under reduced frame budgets.

Position information is essential for language modeling. In softmax transformers, Rotary Position Embeddings (\textit{RoPE}) encode positions through \textit{fixed-angle} rotations, while in linear transformers, order is handled via input-dependent (selective) gating that decays past key-value associations. Selectivity has generally been shown to improve language-related tasks. Inspired by this, we introduce \textit{Selective RoPE}, an \textit{input-dependent} rotary embedding mechanism, that generalizes \textit{RoPE}, and enables rotation in \textit{arbitrary angles} for both linear and softmax transformers. We show that softmax attention already performs a hidden form of these rotations on query-key pairs, uncovering an implicit positional structure. We further show that in state-space models and gated linear transformers, the real part manages forgetting while the imaginary part encodes positions through rotations. We validate our method by equipping gated transformers with \textit{Selective RoPE}, demonstrating that its input-dependent rotations improve performance in language modeling and on difficult sequence tasks like copying, state tracking, and retrieval.

Dyslexia, a neurodevelopmental disorder characterized by persistent reading difficulties, is often linked to reduced activity of the visual word form area in the ventral occipito-temporal cortex. Traditional approaches to studying dyslexia, such as behavioral and neuroimaging methods, have provided valuable insights but remain limited in their ability to test causal hypotheses about the underlying mechanisms of reading impairments. In this study, we use large-scale vision-language models (VLMs) to simulate dyslexia by functionally identifying and perturbing artificial analogues of word processing. Using stimuli from cognitive neuroscience, we identify visual-word-form-selective (VWF-selective) units within VLMs and demonstrate that targeted ablation of these units, unlike ablation of random units, leads to selective impairments in reading tasks while general visual and language comprehension abilities remain intact. In particular, the resulting model matches dyslexic humans' phonological deficits without a significant change in orthographic processing. Additionally, the model’s VWF-selective units predict human-VWFA neural responses better than random units and the ablated model mirrors dyslexic behavior in font sensitivity. Taken together, our modeling results replicate key characteristics of dyslexia and establish a computational framework for investigating reading disorders.

The development of foundation models for functional magnetic resonance imaging (fMRI) time series holds significant promise for predicting phenotypes related to disease and cognition. Current models, however, are often trained using a mask-and-reconstruct objective on small brain regions. This focus on low-level information leads to representations that are sensitive to noise and temporal fluctuations, necessitating extensive fine-tuning for downstream tasks. We introduce Brain-Semantoks, a self-supervised framework designed specifically to learn abstract representations of brain dynamics. Its architecture is built on two core innovations: a semantic tokenizer that aggregates noisy regional signals into robust tokens representing functional networks, and a self-distillation objective that enforces representational stability across time. We show that this objective is stabilized through a novel training curriculum, ensuring the model robustly learns meaningful features from low signal-to-noise time series. We demonstrate that learned representations enable strong performance on a variety of downstream tasks even when only using a linear probe. Furthermore, we provide comprehensive scaling analyses indicating more unlabeled data reliably results in out-of-distribution performance gains without domain adaptation.

Linear autoencoders (LAEs) have gained increasing popularity in recommender systems due to their simplicity and strong empirical performance. Most LAE models, including the Emphasized Denoising Linear Autoencoder (EDLAE) introduced by (Steck, 2020), use quadratic loss during training. However, the original EDLAE only provides closed-form solutions for the hyperparameter choice $b = 0$, which limits its capacity. In this work, we generalize EDLAE objective function into a Decoupled Expected Quadratic Loss (DEQL). We show that DEQL simplifies the process of deriving EDLAE solutions and reveals solutions in a broader hyperparameter range $b > 0$, which were not derived in Steck’s original paper. Additionally, we propose an efficient algorithm based on Miller’s matrix inverse theorem to ensure the computational tractability for the $b > 0$ case. Empirical results on benchmark datasets show that the $b > 0$ solutions provided by DEQL outperform the $b = 0$ EDLAE baseline, demonstrating that DEQL expands the solution space and enables the discovery of models with better testing performance.

Object detection systems are essential in safety-critical applications, but they are vulnerable to object disappearance (OD) threat, in which valid objects become undetected under small input perturbations, creating serious risks. This paper addresses the problem of verifying the robustness of YOLO networks against OD by proposing a three-step probabilistic verification framework: (1) estimating output ranges under a distribution of input perturbations, (2) formally verifying the Non-Maximum Suppression (NMS) process within these ranges, and (3) iteratively refining the results to reduce over-approximation. The framework scales to practical YOLO models. Both theoretical analysis and experimental results demonstrate that our method achieves comparable probabilistic guarantees and provides tighter Intersection-over-Union (IoU) lower bounds while requiring significantly fewer samples than existing methods.

Denoising-based generative models have been significantly advanced by representation-alignment (REPA) loss, which leverages pre-trained visual encoders to guide intermediate network features. However, REPA's reliance on external visual encoders introduces two critical challenges: potential \textit{distribution mismatches} between the encoder's training data and the generation target, and the high \textit{computational costs} of pre-training. Inspired by the observation that REPA primarily aids early layers in capturing robust semantics, we propose an unsupervised alternative that avoids external visual encoder and the assumption of consistent data distribution. We introduce \textit{\textbf{DU}al-\textbf{P}ath condition \textbf{A}lignment} (\textbf{DUPA}), a novel self-alignment framework, which independently noises an image multiple times and processes these noisy latents through decoupled diffusion transformer, then aligns the derived conditions\textemdash low-frequency semantic features extracted from each path. Experiments demonstrate that DUPA achieves FID$=$1.46 on ImageNet 256$\times$256 with only 400 training epochs, outperforming all methods that do not rely on external supervision. Critically, DUPA accelerates training of its base model by 5$\times$ and inference by 10$\times$. DUPA is also model-agnostic and can be readily applied to any denoising-based generative model, showcasing its excellent scalability and generalizability.

There is a growing interest in Universal Multimodal Embeddings (UME), where models are required to generate task-specific representations. While recent studies show that Multimodal Large Language Models (MLLMs) perform well on such tasks, they treat MLLMs solely as encoders, overlooking their generative capacity. However, such an encoding paradigm becomes less effective as instructions become more complex and require compositional reasoning. Inspired by the proven effectiveness of chain-of-thought reasoning, we propose a general Think-Then-Embed (TTE) framework for UME, composed of a reasoner and an embedder. The reasoner MLLM first generates reasoning traces that explain complex queries, followed by an embedder that produces representations conditioned on both the original query and the intermediate reasoning. This explicit reasoning step enables more nuanced understanding of complex multimodal instructions. Our contributions are threefold. First, by leveraging a powerful MLLM reasoner, we achieve state-of-the-art performance on the MMEB-V2 benchmark, surpassing proprietary models trained on massive in-house datasets. Second, to reduce the dependency on large MLLM reasoners, we finetune a smaller MLLM reasoner using high-quality embedding-centric reasoning traces, achieving the best performance among open-source models with a 7% absolute gain over recently proposed models. Third, we investigate strategies for integrating the reasoner and embedder into a unified model for improved efficiency without sacrificing performance.

The Hybrid Online Learning Problem, where features are drawn i.i.d. from an unknown distribution but labels are generated adversarially, is a well-motivated setting positioned between statistical and fully-adversarial online learning. Prior work has presented a dichotomy: algorithms that are statistically-optimal, but computationally intractable \citep{wu2023expected}, and algorithms that are computationally-efficient (given an ERM oracle), but statistically-suboptimal \citep{pmlr-v247-wu24a}. This paper takes a significant step towards achieving statistical optimality and computational efficiency \emph{simultaneously} in the Hybrid Learning setting. To do so, we consider a structured setting, where the Adversary is constrained to pick labels from an expressive, but fixed, class of functions $\mathcal{R}$. Our main result is a new learning algorithm, which runs efficiently given an ERM oracle and obtains regret scaling with the Rademacher complexity of a class derived from the Learner's hypothesis class $\mathcal{H}$ and the Adversary's label class $\mathcal{R}$. As a key corollary, we give an oracle-efficient algorithm for computing equilibria in stochastic zero-sum games when action sets may be high-dimensional but the payoff function exhibits a type of low-dimensional structure. Technically, we develop a number of novel tools for the design and analysis of our learning algorithm, including a novel Frank-Wolfe reduction with "truncated entropy regularizer" and a new tail bound for sums of "hybrid'' martingale difference sequences.

Learning the dynamics of a process given sampled observations at several time points is an important but difficult task in many scientific applications. When no ground-truth trajectories are available, but one has only snapshots of data taken at discrete time steps, the problem of modelling the dynamics, and thus inferring the underlying trajectories, can be solved by multi-marginal generalisations of flow matching algorithms. This paper proposes a novel flow matching method that overcomes the limitations of existing multi-marginal trajectory inference algorithms. Our proposed method, ALI-CFM, uses a GAN-inspired adversarial loss to fit neurally parametrised interpolant curves between source and target points such that the marginal distributions at intermediate time points are close to the observed distributions. The resulting interpolants are smooth trajectories that, as we show, are unique under mild assumptions. These interpolants are subsequently marginalised by a flow matching algorithm, yielding a trained vector field for the underlying dynamics. We showcase the versatility and scalability of our method by outperforming the existing baselines on spatial transcriptomics and cell tracking datasets, while performing on par with them on single-cell trajectory prediction.

Optimizers are critical for training deep neural networks. Existing training processes rely on a single static optimizer (e.g., SGD) or a simple hybrid of two optimizers, which miss the opportunity to exploit evolving dynamics in different training states, degrading model quality and convergence. In this paper, we propose a novel dynamic optimizer switching method called **D**ynamic **O**ptimizer **I**nterleaving **T**raining (DOIT) method, which builds surrogate models to predict different optimizers' performance from current parameter states. DOIT uses an acquisition function that combines the results from surrogate models with transferability assessments and process information to select a suitable optimizer for the subsequent training. Experiments on various models and tasks (e.g., image and text classification, machine translation, and object detection) show that DOIT effectively enhances the training, achieving faster convergence (i.e., 2\% to 10\% faster) and higher accuracy (i.e., 1\% to 3\% improvement). Additional independent experiments and case studies further validate DOIT's effectiveness.

Sheaf neural networks (SNNs) leverage cellular sheaves to induce flexible diffusion processes on graphs, generalizing the diffusion mechanism of classical graph neural networks. While SNNs have been shown to cope well with heterophilic tasks and alleviate oversmoothing, we argue that there is further room for improving sheaf diffusion. More specifically, we show that SNNs do not allow nodes to independently choose how they cooperate with their neighbors, i.e., whether they convey and/or gather information to/from their neighbors. To address this issue, we first introduce the notion of cellular sheaves over directed graphs and characterize their in- and out-degree Laplacians. We then leverage our construction to propose Cooperative Sheaf Neural Network (CSNN). Additionally, we formally characterize its receptive field and prove that it allows nodes to selectively attend (listen) to arbitrarily far nodes while ignoring all others in their path, which is key to alleviating oversquashing. Our results on synthetic data empirically substantiate our claims, showing that CSNN can handle long-range interactions while avoiding oversquashing. We also show that CSNN performs strongly in heterophilic node classification and long-range graph classification benchmarks.

The growing demand for prenatal ultrasound imaging has intensified a global shortage of trained sonographers, creating barriers to essential fetal health monitoring. Deep learning has the potential to enhance sonographers' efficiency and support the training of new practitioners. Vision-Language Models (VLMs) are particularly promising for ultrasound interpretation, as they can jointly process images and text to perform multiple clinical tasks within a single framework. However, despite the expansion of VLMs, no standardized benchmark exists to evaluate their performance in fetal ultrasound imaging. This gap is primarily due to the modality’s challenging nature, operator dependency, and the limited public availability of datasets. To address this gap, we present Fetal-Gauge, the first and largest visual question answering benchmark specifically designed to evaluate VLMs across various fetal ultrasound tasks. Our benchmark comprises over 42,000 images and 93,000 question-answer pairs, spanning anatomical plane identification, visual grounding of anatomical structures, fetal orientation assessment, clinical view conformity, and clinical diagnosis. We systematically evaluate several state-of-the-art VLMs, including general-purpose and medical-specific models, and reveal a substantial performance gap: the best-performing model achieves only 55\% accuracy, far below clinical requirements. Our analysis identifies critical limitations of current VLMs in fetal ultrasound interpretation, highlighting the urgent need for domain-adapted architectures and specialized training approaches. Fetal-Gauge establishes a rigorous foundation for advancing multimodal deep learning in prenatal care and provides a pathway toward addressing global healthcare accessibility challenges. Our benchmark is publicly available at www.github.com

Quantization and cache mechanisms are typically applied individually in efficient generation tasks, each showing notable potential for acceleration. However, their joint effect on efficiency remains under-explored. Through both empirical investigation and theoretical analysis, we find that that combining quantization with caching is non-trivial, as it introduces two major challenges that severely degrade performance: (i) the sample efficacy of calibration datasets in post-training quantization (PTQ) is significantly eliminated by cache operation; (ii) the joint use of the two mechanisms exacerbates exposure bias in the sampling distribution, leading to amplified error accumulation during generation. In this work, we take advantage of these two acceleration mechanisms and propose a hybrid acceleration method by tackling the above challenges, aiming to further improve the efficiency of tasks while maintaining excellent generation capability. Concretely, a temporal-aware parallel clustering (TAP) is designed to dynamically improve the sample selection efficacy for the calibration within PTQ for different diffusion steps. A variance compensation (VC) strategy is derived to correct the sampling distribution. It mitigates exposure bias through an adaptive correction factor generation. Extensive experiments demonstrate that our method is broadly applicable to diverse generation tasks, achieving up to 12.7x acceleration while preserving competitive generation quality.

Monitoring large language models' (LLMs) activations is an effective way to detect harmful requests before they lead to unsafe outputs. However, traditional safety monitors often require the same amount of compute for every query. This creates a trade-off: expensive monitors waste resources on easy inputs, while cheap ones risk missing subtle cases. We argue that safety monitors should be flexible--costs should rise only when inputs are difficult to assess, or when more compute is available. To achieve this, we introduce Truncated Polynomial Classifiers (TPCs), a natural extension of linear probes for dynamic activation monitoring. Our key insight is that polynomials can be trained and evaluated progressively, term-by-term. At test-time, one can early-stop for lightweight monitoring, or use more terms for stronger guardrails when needed. TPCs provide two modes of use. First, as a safety dial: by evaluating more terms, developers and regulators can "buy" stronger guardrails from the same model. Second, as an adaptive cascade: clear cases exit early after low-order checks, and higher-order guardrails are evaluated only for ambiguous inputs, reducing overall monitoring costs. On two large-scale safety datasets (WildGuardMix and BeaverTails), for 4 models with up to 30B parameters, we show that TPCs compete with or outperform MLP-based probe baselines of the same size, all the while being more interpretable than their black-box counterparts. Our anonymous code is available at https://anonymous.4open.science/r/tpc-anon-0708.

Autoregressive transformers are the backbone of modern large language models. Despite their success, inference remains slow due to strictly sequential prediction. Prior attempts to predict multiple tokens per step typically impose independence assumptions across tokens, which limits their ability to match the full expressiveness of standard autoregressive models. In this work, we break this paradigm by proposing an efficient and universal framework to jointly predict multiple tokens in a single transformer call, without limiting the representational power. Inspired by ideas from inverse autoregressive normalizing flows, we convert a series of random variables deterministically into a token sequence, incorporating the sampling procedure into a trained model. This allows us to train parallelized models both from scratch and by distilling an existing autoregressive model. Empirically, our distilled model matches its teacher's output for an average of close to 50 tokens on toy data and 5 tokens on a coding dataset, all within a single forward pass.