ICLR 2026 论文列表

Unlocking advanced reasoning in large language model agents is hindered by a scarcity of training data situated at the very frontier of their capabilities. We address this with a novel data synthesis approach inspired by the educational theory of the Zone of Proximal Development (ZPD), which conceptualizes this frontier as tasks an LLM cannot solve independently but can master with guidance. We operationalize this principle through the AgentFrontier Data Engine, an automated pipeline that synthesizes high-quality, multidisciplinary data situated precisely within an LLM's ZPD. The engine yields two synergistic outputs: knowledge-intensive data for continued pre-training and frontier-level reasoning trajectories for post-training. Concurrently, it produces the ZPD Exam, a self-evolving benchmark for evaluating agent capabilities by compelling them to reason beyond their parameterized knowledge. By training our AgentFrontier-30B-A3B model on the synthesized data, we achieve state-of-the-art results on demanding benchmarks like Humanity's Last Exam, outperforming several leading proprietary agents. This work establishes ZPD-guided data synthesis as a scalable and effective paradigm for cultivating increasingly capable LLM agents.

Attention sinks and compression valleys have attracted significant attention as two puzzling phenomena in large language models, but have been studied in isolation. In this work, we present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream. We prove theoretically that massive activations necessarily produce representational compression and establish bounds on the resulting entropy reduction. Through experiments across several models (410M--120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys and attention sinks emerge simultaneously. Targeted ablation validates our theoretical predictions. This unified view motivates us to propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations. Specifically, we posit that Transformer-based LLMs process tokens in three distinct phases: (1) broad mixing in the early layers, (2) compressed computation with limited mixing in the middle layers, and (3) selective refinement in the late layers. Our framework helps explain why embedding tasks perform best at intermediate layers, whereas generation tasks benefit from full-depth processing, clarifying differences in task-dependent representations.

We introduce SeerAttention-R, a sparse attention framework specifically tailored for the long decoding of reasoning models. Extended from SeerAttention, SeerAttention-R retains the design of learning attention sparsity through a self-distilled gating mechanism, while removing query pooling to accommodate auto-regressive decoding. With a lightweight plug-in gating, SeerAttention-R is flexible and can be easily integrated into existing pretrained model without modifying the original parameters. We demonstrate that SeerAttention-R, trained on just 0.4B tokens, maintains near-lossless reasoning accuracy with 4K token budget in AIME benchmark under large sparse attention block sizes (64/128). Using TileLang, we develop a highly optimized sparse decoding kernel that achieves near-theoretical speedups of up to 9x over FlashAttention-3 on H100 GPU at 90\% sparsity.

Graph Neural Networks (GNNs) leverage the graph structure to transmit information between nodes, typically through the message-passing mechanism. While these models have found a wide variety of applications, they are known to suffer from over-squashing, where information from a large receptive field of node representations is collapsed into a single fixed sized vector, resulting in an information bottleneck. In this paper, we re-examine the over-squashing phenomenon through the lens of model storage and retrieval capacity, which we define as the amount of information that can be stored in a node’s representation for later use. We study some of the limitations of existing tasks used to measure over-squashing and introduce a new synthetic task to demonstrate that an information bottleneck can saturate this capacity. Furthermore, we adapt ideas from the sequence modeling literature on associative memories, fast weight programmers, and the xLSTM model to develop a novel GNN architecture with improved capacity. We demonstrate strong performance of this architecture both on our capacity synthetic task, as well as a range of real-world graph benchmarks.

Recent research has introduced distributed self-supervised learning (D-SSL) approaches to leverage vast amounts of unlabeled decentralized data. However, D-SSL faces the critical challenge of data heterogeneity, and there is limited theoretical understanding of how different D-SSL frameworks respond to this challenge. To fill this gap, we present a rigorous theoretical analysis of the robustness of D-SSL frameworks under non-IID (non-independent and identically distributed) settings. Our results show that pre-training with Masked Image Modeling (MIM) is inherently more robust to heterogeneous data than Contrastive Learning (CL), and that the robustness of decentralized SSL increases with average network connectivity, implying that federated learning (FL) is no less robust than decentralized learning (DecL). These findings provide a solid theoretical foundation for guiding the design of future D-SSL algorithms. To further illustrate the practical implications of our theory, we introduce MAR loss, a refinement of the MIM objective with local-to-global alignment regularization. Extensive experiments across model architectures and distributed settings validate our theoretical insights, and additionally confirm the effectiveness of MAR loss as an application of our analysis.

Generalization—the ability to perform well beyond the training context—is a hallmark of biological and artificial intelligence, yet anticipating unseen failures remains a central challenge. Conventional approaches often take a bottom-up mechanistic route by reverse-engineering interpretable features or circuits to build explanatory models. However, they provide little top-down guidance such as system-level measurements that predict and prevent failures. Here we propose a complementary diagnostic paradigm for studying generalization failures. Rather than mapping out detailed internal mechanisms, we use task-relevant measures to probe structure–function links, identify prognostic indicators, and test predictions in real-world settings. In image classification, we find that task-relevant geometric properties of in-distribution (ID) object manifolds consistently signal poor out-of-distribution (OOD) generalization. In particular, reductions in two geometric measures—effective manifold dimensionality and utility—predict weaker OOD performance across diverse architectures, optimizers, and datasets. We apply this finding to transfer learning with ImageNet-pretrained models, each available with multiple weight variants. We consistently find that the same geometric patterns predict OOD transfer performance more reliably than ID accuracy. This work demonstrates that representational geometry can expose hidden vulnerabilities, offering more robust guidance for model selection.

Recent Text-to-Video (T2V) models have demonstrated powerful capability in visual simulation of real-world geometry and physical laws, indicating its potential as implicit world models. Inspired by this, we explore the feasibility of leveraging the video generation prior for viewpoint planning from given 4D scenes, since videos internally accompany dynamic scenes with natural viewpoints. To this end, we propose a two-stage paradigm to adapt pre-trained T2V models for viewpoint prediction, in a compatible manner. First, we inject the 4D scene representation into the pre-trained T2V model via an adaptive learning branch, where the 4D scene is viewpoint-agnostic and the conditional generated video embeds the viewpoints visually. Then, we formulate viewpoint extraction as a hybrid-condition guided camera extrinsic denoising process. Specifically, a camera extrinsic diffusion branch is further introduced onto the pre-trained T2V model, by taking the generated video and 4D scene as input. Experimental results show the superiority of our proposed method over existing competitors, and ablation studies validate the effectiveness of our key technical designs. To some extent, this work proves the potential of video generation models toward 4D interaction in real world.

Jailbreaking refers to techniques that bypass the safety alignment of large language models (LLMs) to elicit harmful outputs, and automated red-teaming has become a key approach for detecting such vulnerabilities before deployment. However, most existing red-teaming methods operate directly in text space, where they tend to generate semantically similar prompts and thus fail to probe the broader spectrum of latent vulnerabilities within a model. To address this limitation, we shift the exploration of jailbreak strategies from conventional text space to the model’s latent activation space and propose STAR (**ST**rategy-driven **A**utomatic Jailbreak **R**ed-teaming), a black-box framework for systematically generating jailbreak prompts. STAR is composed of two modules: (i) strategy generation module, which extracts the principal components of existing strategies and recombines them to generate novel ones; and (ii) prompt generation module, which translates abstract strategies into concrete jailbreak prompts with high success rates. Experimental results show that STAR substantially outperforms state-of-the-art baselines in terms of both attack success rate and strategy diversity. These findings highlight critical vulnerabilities in current alignment techniques and establish STAR as a more powerful paradigm for comprehensive LLM security evaluation.

Deep stereo disparity estimation has long been dominated by a \textbf{matching-centric paradigm}, built on constructing cost volumes and iteratively refining local correspondences. Despite its success, this paradigm exhibits an intrinsic vulnerability: visual ambiguities from occlusion or non-Lambertian surfaces invevitably induce errorneous matches that refinement cannot recover. This paper introduces \textbf{DispViT}, a new architecture that establishes a \textbf{regression-centric paradigm}. Instead of explicit matching, DispViT directly regresses disparity from tokenized binocular representations using a single-stream Vision Transformer. This is enabled by a set of lightweight yet critical designs, such as a probability-based disparity parameterization for stable training and an asymmetrically initialized stereo tokenizer for effective view distinction. To better align the two views during stereo tokenization, we introduce a novel shift-embedding mechanism that encodes different disparity shifts into channel groups, preserving geometric cues even under large view displacements. A lightweight refinement module then sharpens the regressed disparity map for fine-grained accuracy. By prioritizing holistic regression over explicit matching, DispViT streamlines the stereo pipeline while improving robustness and efficiency. Experiments on standard benchmarks show that our approach achieves state-of-the-art accuracy, with strong resilience to matching ambiguities and wide disparity ranges. Code will be released.

Cyclic learning, which involves training with pairs of inverse tasks and utilizes cycle-consistency in the design of loss functions, has emerged as a powerful paradigm for weakly-supervised learning. However, its potential remains under-explored due to the current methods’ narrow focus on domain-specific implementations. In this work, we develop generalized solutions for both pairwise cycle-consistent tasks and self-cycle-consistent tasks. By formulating cross-domain mappings as conditional probability functions, we reformulate the cycle-consistency objective as an evidence lower bound optimization problem via variational inference. Based on this formulation, we further propose two training strategies for arbitrary cyclic learning tasks: single-step optimization and alternating optimization. Our framework demonstrates broad applicability across diverse tasks. In unpaired image translation, it not only provides a theoretical justification for CycleGAN but also leads to CycleGN—a competitive GAN-free alternative. For unsupervised tracking, CycleTrack and CycleTrack-EM achieve state-of-the-art performance on multiple benchmarks. This work establishes the theoretical foundations of cyclic learning and offers a general paradigm for future research.

Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for complex reasoning tasks with clear correctness signals such as math and coding. However, extending it to real-world reasoning tasks is challenging, as evaluation depends on nuanced, multi-criteria judgments rather than binary correctness. Instance-specific rubrics have recently been used in evaluation benchmarks to capture such judgments, but their potential as reward signals for on-policy post-training remains underexplored. We introduce $\textbf{Rubrics as Rewards (\textit{RaR})}$, an on-policy reinforcement learning method that extends RLVR beyond verifiable domains by using rubric-based feedback. Across both medical and science domains, we evaluate multiple strategies for aggregating rubric feedback into rewards. The best RaR variant achieves relative improvements of up to 31\% on HealthBench and 7\% on GPQA-Diamond over popular LLM-as-judge baselines that rely on direct Likert-based rewards. These results demonstrate that RaR-trained policies adapt well to diverse evaluation formats, performing strongly on both rubric-based and multiple-choice tasks. Moreover, we find that using rubrics as structured reward signals yields better alignment for smaller judges and reduces performance variance across judge scales.

We present NOVA3R, an effective approach for non-pixel-aligned 3D reconstruction from a set of unposed images, in a feed-forward manner. Unlike pixel-aligned methods that tie geometry to per-ray predictions, our formulation learns a global, view-agnostic scene representation that decouples reconstruction from pixel alignment. This addresses two key limitations in pixel-aligned 3D: (1) it recovers both visible and invisible regions with a complete scene representation, and (2) it produces physically plausible geometry with fewer duplicated structures in overlapping regions. To achieve this, we introduce a scene-token mechanism that aggregates information across unposed images and a diffusion-based 3D decoder that reconstructs complete, non-pixel-aligned point clouds. Extensive experiments on both scene-level and object-level datasets demonstrate that NOVA3R outperforms state-of-the-art methods in terms of reconstruction accuracy and completeness.

Deep neural networks have been increasingly used in safety-critical applications such as medical diagnosis and autonomous driving. However, many studies suggest that they are prone to being poorly calibrated and have a propensity for overconfidence, which may have disastrous consequences. In this paper, unlike standard training such as stochastic gradient descent, we show that the recently proposed sharpness-aware minimization (SAM) counteracts this tendency towards overconfidence. The theoretical analysis suggests that SAM allows us to learn models that are already well-calibrated by implicitly maximizing the entropy of the predictive distribution. Inspired by this finding, we further propose a variant of SAM, coined as CSAM, to ameliorate model calibration. Extensive experiments on various datasets, including ImageNet-1K, demonstrate the benefits of SAM in reducing calibration error. Meanwhile, CSAM performs even better than SAM and consistently achieves lower calibration error than other approaches.

As large language models continue to scale, their growing computational and storage demands pose significant challenges for real-world deployment. In this work, we investigate redundancy within Transformer-based models and propose an entropy-based pruning strategy to enhance efficiency while maintaining performance. Empirical analysis reveals that the entropy of hidden representations decreases in the early blocks but progressively increases across most subsequent blocks. This trend suggests that entropy serves as a more effective measure of information richness within computation blocks. Unlike cosine similarity, which primarily captures geometric relationships, entropy directly quantifies uncertainty and information content, making it a more reliable criterion for pruning. Extensive experiments demonstrate that our entropy-based pruning approach surpasses cosine similarity-based methods in reducing model size while preserving accuracy, offering a promising direction for efficient model deployment.

Fine-tuning Large Language Models (LLMs) with downstream data is often considered time-consuming and expensive. Structured pruning methods are primarily employed to improve the inference efficiency of pre-trained models. Meanwhile, they often require additional time and memory for training, knowledge distillation, structure search, and other strategies, making efficient model fine-tuning challenging to achieve. To simultaneously enhance the training and inference efficiency of downstream task fine-tuning, we introduce GradPruner, which can prune layers of LLMs guided by gradients in the early stages of fine-tuning. GradPruner uses the cumulative gradients of each parameter during the initial phase of fine-tuning to compute the Initial Gradient Information Accumulation Matrix (IGIA-Matrix) to assess the importance of layers and perform pruning. We sparsify the pruned layers based on the IGIA-Matrix and merge them with the remaining layers. Only elements with the same sign are merged to reduce interference from sign variations. We conducted extensive experiments on two LLMs across eight well-known datasets in downstream tasks. Including medical, financial, and general benchmark tasks. The results demonstrate that GradPruner has achieved a parameter reduction of 40% with only a 0.99% decrease in accuracy. Our code is available at https://anonymous.4open.science/r/LLM-GradPrune-436D.

Constant bit-size Transformers are known to be Turing complete, but existing constructions require $\Omega(s(n))$ chain-of-thought (CoT) steps per simulated Turing machine (TM) step, leading to impractical reasoning lengths. In this paper, we significantly reduce this efficiency gap by proving that any $(t(n),s(n))$-bounded multi-tape TM can be simulated by a constant bit-size Transformer with an optimal $O(s(n))$-long context window and only $O(s(n)^c)$ CoT steps per TM step, where $c>0$ can be made arbitrarily small by letting the Transformers' head-layer product sufficiently large. In addition, our construction shows that sparse attention with fixed geometric offsets suffices for efficient universal computation. Our proof leverages multi-queue TMs as a bridge. The main technical novelty is a more efficient simulation of multi-tape TMs by synchronous multi-queue TMs, improving both time and space complexity under stricter model assumptions.

We study offline change point localization and inference in dynamic multilayer random dot product graphs (D-MRDPGs), where at each time point, a multilayer network is observed with shared node latent positions and time-varying, layer-specific connectivity patterns. We propose a novel two-stage algorithm that combines seeded binary segmentation with low-rank tensor estimation, and establish its consistency in estimating both the number and locations of change points. Furthermore, we derive the limiting distributions of the refined estimators under both vanishing and non-vanishing jump regimes. To the best of our knowledge, this is the first result of its kind in the context of dynamic network data. We also develop a fully data-driven procedure for constructing confidence intervals. Extensive numerical experiments demonstrate the superior performance and practical utility of our methods compared to existing alternatives.

High dynamic range (HDR) video can be reconstructed from low dynamic range (LDR) sequences with alternating exposures. However, most existing methods overlook the degradations (e.g., noise and blur) in LDR frames, focusing only on the brightness and position differences between them. To address this gap, we propose DeAltHDR, a novel framework for high-quality HDR video reconstruction from degraded sequences. Our framework addresses two key challenges. First, noisy and blurry contents complicate inter-frame alignment. To tackle this, we propose a flow-guided masked attention that leverages optical flow for a dynamic sparse cross-attention computation, achieving superior performance while maintaining efficiency. Notably, its controllable attention ratio allows for adaptive inference costs. Second, the lack of real-world paired data hinders practical deployment. We overcome this with a two-stage training paradigm: the model is first pre-trained on our newly introduced synthetic paired dataset and subsequently fine-tuned on unlabeled real-world videos via a proposed self-supervised method. Experiments show our method outperforms state-of-the-art ones. The datasets and code will be publicly available.

Kolmogorov-Arnold Networks (KANs) have emerged as a promising alternative to traditional neural networks, offering enhanced interpretability based on the Kolmogorov-Arnold representation theorem. While their empirical success is growing, a theoretical understanding of their training dynamics remains nascent. This paper investigates the optimization of a two-layer KAN in the overparameterized regime, focusing on a simplified yet insightful setting where only the first-layer coefficients are trained via gradient descent. Our main result establishes that, provided the network is sufficiently wide, this training method is guaranteed to converge to a global minimum and achieve zero training error. Furthermore, we derive a novel, fine-grained convergence rate that explicitly connects the optimization speed to the structure of the data labels through the eigenspectrum of the KAN Tangent Kernel (KAN-TK). Our analysis reveals a key advantage of this architecture: guaranteed convergence is achieved with a hidden layer width of $m=\mathcal{O}(n^2)$, a significant polynomial improvement over the $m=\mathcal{O}(n^6)$ requirement for classic two-layer neural networks using ReLU activation functions and analyzed within the same Tangent Kernel framework. We validate our theoretical findings with numerical experiments that corroborate our predictions on convergence speed and the impact of label structure.