ICLR 2026 Papers

Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in visual understanding and reasoning, but they also impose significant computational burdens due to long visual sequence inputs. Recent works address this issue by pruning unimportant visual tokens, achieving substantial computational reduction while maintaining model performance. The core of token pruning lies in determining token importance, with current approaches primarily relying on attention scores from vision encoders or Large Language Models (LLMs). In this paper, we analyze the effectiveness of attention mechanisms in both vision encoders and LLMs. We find that vision encoders suffer from attention sink, leading to poor focus on informative foreground regions, while in LLMs, although prior studies have identified attention bias toward token positions, text-to-vision attention demonstrates resistance to this bias and enables effective pruning guidance in middle layers. Based on these observations, we propose $\textbf{LearnPruner}$, a two-stage token pruning framework that first removes redundant vision tokens via a learnable pruning module after the vision encoder, then retains only task-relevant tokens in the LLM's middle layer. Experimental results show that our LearnPruner can preserve approximately 95\% of the original performance while using only 5.5\% of vision tokens, and achieve 3.2$\times$ inference acceleration, demonstrating a superior accuracy-efficiency trade-off.

Despite rapid advancements in large language models (LLMs), the token-level autoregressive nature constrains their complex reasoning capabilities. To enhance LLM reasoning, inference-time techniques, including Chain/Tree/Graph-of-Thought(s), successfully improve the performance, as they are fairly cost-effective by guiding reasoning through external logical structures without modifying LLMs' parameters. However, these manually predefined, task-agnostic frameworks are applied uniformly across diverse tasks, lacking adaptability. To improve this, we propose **RL-of-Thoughts (RLoT)**, where we train a lightweight navigator model with reinforcement learning (RL) to generate task-adaptive logical structures at inference time, enhancing LLM reasoning. Specifically, we design five basic logic blocks from the perspective of human cognition. During the reasoning process, the trained RL navigator dynamically selects the suitable logic blocks and combines them into task-specific logical structures according to problem characteristics. Experiments across multiple reasoning benchmarks (AIME, MATH, GPQA, etc.) with multiple LLMs (GPT, Llama, Qwen, and DeepSeek) illustrate that RLoT outperforms established inference-time techniques in most cases and improves up to 13.4% in challenging situations. Remarkably, with less than 3K parameters, our RL navigator is able to make sub-10B LLMs comparable to 100B-scale counterparts. Moreover, the RL navigator demonstrates strong transferability: a model trained on one specific LLM-task pair can effectively generalize to unseen LLMs and tasks. Our code is open-source at https://anonymous.4open.science/r/RL-LLM-Reasoning-1A30.

Accurate detection of clinically meaningful events in healthcare time-series data is crucial for reliable downstream analysis and decision support. However, most existing methods struggle to jointly localize event boundaries and classify event types; even detection transformer (DETR)-based approaches show limited performance when confronted with extremely sparse events typical of clinical recordings. To address these challenges, we propose a coarse-to-fine detection framework combining a global context explorer, a local detail inspector, and an adaptive gating module (AGM) that fuses multiple label perspectives. The AGM uses transformed labels—encoding event presence and temporal position—to improve learning on sparse events. This design acts as a switch that selectively activates detailed feature extraction only when an event is likely, thereby reducing noise and improving efficiency in sparse settings. We evaluate our framework on diverse healthcare datasets—including arrhythmia detection, emotion recognition, and human-activity monitoring—and demonstrate substantial performance gains over existing DETR-based models, with particularly strong improvements in sparse event detection. With precise and robust event detection, our framework enables interpretation and actionable insights in real-world clinical applications.

Data-centric distillation, including data augmentation, selection, and mixing, offers a promising path to creating smaller, more efficient student Large Language Models (LLMs) that retain strong reasoning abilities. However, there still lacks a comprehensive benchmark to systematically assess the effect of each distillation approach. This paper introduces DC-CoT, the first data-centric benchmark that investigates data manipulation in chain-of-thought (CoT) distillation from method, model and data perspectives. Utilizing various teacher models (e.g., o4-mini, Gemini-Pro, Claude-3.5) and student architectures (e.g., 3B, 7B parameters), we rigorously evaluate the impact of these data manipulations on student model performance across multiple reasoning datasets, with a focus on in-distribution (IID) and out-of-distribution (OOD) generalization, and cross-domain transfer. Our findings aim to provide actionable insights and establish best practices for optimizing CoT distillation through data-centric techniques, ultimately facilitating the development of more accessible and capable reasoning models. The nonymous codebase can be accessed https://anonymous.4open.science/r/DC-COT-FF4C/

Training billion-parameter Transformers is often brittle, with transient loss spikes and divergence that waste compute. Even though the recently developed Edge of Stability (EoS) theory provides a powerful tool to understand and control the stability of optimization methods via the (preconditioned) curvature, these curvature-controlling methods are not popular in large-scale Transformer training due to the complexity of curvature estimation. To this end, we first introduce a fast online estimator of the largest (preconditioned) Hessian eigenvalue (i.e., curvature) based on a warm-started variant for power iteration with Hessian–vector products. We show theoretically, and verify empirically, that the proposed method makes per-iteration curvature tracking feasible at billion-parameter scale while being more accurate. Using this tool, we find that training instabilities coincide with surges in preconditioned curvature and that curvature grows with depth. Motivated by these observations, we propose architecture warm-up: progressively growing network depth to carefully control the preconditioned Hessian and stabilize training. Experiments on large Transformers validate that our approach enables efficient curvature tracking and reduces instabilities compared to existing state-of-the-art stabilization techniques without slowing down convergence.

Causal attention has become a foundational mechanism in autoregressive Vision-Language models (VLMs), unifying textual and visual inputs under a single generative framework. However, existing causal mask-based strategies are inherited from large language models (LLMs) where they are tailored for text-only decoding, and their adaptation to vision tokens is insufficiently addressed in the prefill stage. Strictly masking future positions for vision queries introduces overly rigid constraints, which hinder the model’s ability to leverage future context that often contains essential semantic cues for accurate inference. In this work, we empirically investigate how different causal masking strategies affect vision-language inference and then propose a family of future-aware attentions tailored for this setting. We first empirically analyze the effect of previewing future tokens for vision queries and demonstrate that rigid masking undermines the model’s capacity to capture useful contextual semantic representations. Based on these findings, we propose a lightweight attention family that aggregates future visual context into past representations via pooling, effectively preserving the autoregressive structure while enhancing cross-token dependencies. We evaluate a range of causal masks across diverse vision-language inference settings and show that selectively compressing future semantic context into past representations benefits the inference.

Real-world out-of-distribution (OOD) data exhibit broad, continually evolving distributions, rendering reliance solely on in-distribution (ID) data insufficient for robust detection. Consequently, methods leveraging auxiliary Outlier Exposure (OE) data have emerged, substantially enhancing generalization by jointly fine-tuning models on ID and large-scale OE data. However, many existing approaches primarily enforce orthogonality between ID and OE features while pushing OE predictions toward near-uniform, low-confidence scores, thus overlooking the controllability of representation geometry. We propose Vast Predefined Classifiers (VPC), which constructs a pre-specified Orthogonal Equiangular Feature Space (OEFS) to explicitly separate ID and OOD representations while capturing the rich variability of OOD features. We employ evidential priors to align ID features with their class-specific Equiangular Basic Vectors (EBVs), thereby preserving ID performance. In parallel, a new VEBV loss encourages OE features to explore the subspace spanned by Vast EBVs (VEBVs), enabling a rich characterization of diverse OOD patterns. This dual optimization, coupled with the prescribed geometric representation space, promotes optimal orthogonality between ID and OOD representations. Furthermore, we introduce the VPC Score, a discriminative metric based on the L2 activation intensity of features over the predefined classifiers. Extensive experiments across diverse OOD settings and training paradigms on benchmarks including CIFAR-10/100 and the ImageNet-1k, demonstrate strong and robust performance, validating VPC's effectiveness.

Unlike vision and language domains, graph learning lacks a shared input space, as input features differ across graph datasets not only in semantics, but also in value ranges and dimensionality. This misalignment prevents graph models from generalizing across datasets, limiting their use as foundation models. In this work, we propose ALL-IN, a simple and theoretically grounded method that enables transferability across datasets with different input features. Our approach projects node features into a shared random space and constructs representations via covariance-based statistics, thus eliminating dependence on the original feature space. We show that the computed node-covariance operators and the resulting node representations are invariant in distribution to permutations of the input features. We further demonstrate that the expected operator exhibits invariance to general orthogonal transformations of the input features. Empirically, ALL-IN achieves strong performance across diverse node- and graph-level tasks on unseen datasets with new input features, without requiring architecture changes or retraining. These results point to a promising direction for input-agnostic, transferable graph models.

Aligning Large Language Models (LLMs) with human preferences through finetuning is resource-intensive, motivating lightweight alternatives at test time. We address test-time alignment through the lens of sequential decision making, a perspective that reveals two fundamental challenges. When actions are defined at the token level, as in guided decoding, alignment suffers from the curse of horizon. Conversely, when actions are at the response level, as in traditional iterative refinement, the curse of dimensionality emerges. To resolve this trade-off, we draw inspiration from Model Predictive Control (MPC) in control theory to propose Textual Model Predictive Control (TMPC), a novel predictive planning framework adapted for aligning LLMs at inference time. A key limitation of standard MPC is its reliance on predefined, hard segment boundaries, which are often absent in text generation. TMPC overcomes this by introducing two principles inspired by hierarchical reinforcement learning: (1) Hindsight Subgoal Identification, where TMPC analyzes generation subgoals to retrospectively identify high-reward intermediate outputs as subgoals. This allows the framework to discover meaningful, task-specific planning steps (e.g., a sentence in machine translation or a bug fix in code generation.). (2) Subgoal-Conditioned Re-Generation, where these identified subgoals are used to guide subsequent planning iterations. By conditioning on these proven, high-quality subgoals, TMPC ensures stable improvement by building upon previously validated successes. TMPC is evaluated on three tasks with distinct segmentation properties: discourse-level translation, long-form response generation, and program synthesis. The results demonstrate that TMPC consistently improves performance, highlighting the generality.

Distributed learning has achieved remarkable success in training deep neural networks (DNNs) on large datasets, but the communication bottleneck limits its scalability. Various compression techniques have been proposed to alleviate this limitation; however, they either use fixed parameters throughout training or rely on complex and computationally intensive methods to adapt compression parameters. Instead of the hard-to-tune hyperparameters required by adaptive compressors, this paper investigates the impact of two fundamental factors in DNN training—the layer size of the networks and their training phases—to design a simple yet efficient dynamic scheduler for any compressor, guiding the selection of compression parameters. We present a **L**ightweight **E**fficient **G**r**A**dient **C**ompression strategy**Y** or LEGACY, which, in theory, can work with any compression technique to produce a simple dynamic counterpart. We benchmark LEGACY on distributed and federated training, involving seven different DNN architectures, ranging from ResNet, Transformer-XL, to GPT-2, across large and challenging datasets, including ImageNet, WikiText-103, and OpenWebText. On ImageNet-1K, with an equivalent average data volume, LEGACY's dynamic compression strategies improve the Top-1 accuracy of ResNet-50 by 7-11% compared to uniform Top-0.1% compression, while on WikiText-103, the layer-based dynamic strategy reduces the perplexity of Transformer-XL by ~26% relative to the same baseline. In addition, we evaluate LEGACY under constrained and federated settings, and demonstrate that it scales effectively to a 100-worker configuration while maintaining strong accuracy under aggressive compression. We publish anonymized code at: https://github.com/LEGACY-compression/LEGACY.

Large language models (LLMs) have evolved into agentic systems capable of autonomous tool use and multi-step reasoning for complex problem-solving. However, post-training approaches building upon general-purpose foundation models consistently underperform in agentic tasks, particularly in open-source implementations. We identify the root cause: the absence of robust agentic foundation models forces models during post-training to simultaneously learn diverse agentic behaviors while aligning them to expert demonstrations, thereby creating fundamental optimization tensions. To this end, we are the first to propose incorporating Agentic Continual Pre-training (Agentic CPT) into the deep research agents training pipeline to build powerful agentic foundational models. Based on this approach, we develop a deep research agent model named AgentFounder. We evaluate our AgentFounder-30B on 10 benchmarks and achieve state-of-the-art performance while retains strong tool-use ability, notably 39.9% on BrowseComp-en, 43.3% on BrowseComp-zh, and 31.5% Pass@1 on HLE.

Reinforcement Learning from Verifiable Rewards (RLVR) has significantly improved the performance of large language models (LLMs) on tasks with gold ground truth, such as code generation and mathematical reasoning. However, its application to open-ended question answering (QA) remains challenging, primarily due to the absence of reliable evaluation and verifiable reward signals. This difficulty is further compounded by the limitations of existing evaluation paradigms. Previous approaches typically rely on human feedback or LLM-as-judge strategies, which are costly, prone to reward hacking, and often fail to provide sufficiently discriminative or interpretable evaluation signals. To address these limitations, we introduce a schema for generating case-wise rubrics that are question-specific, content-based and stylistically sensitive, thereby evaluating both factual soundness and writing quality. Building on this schema, we propose QuRL (Open-Ended QA with Rubric-guided Reinforcement Learning), a framework that automatically mines rubrics for each question from easily accessible online sources and leverages them as reward signals. With these rubrics, QuRL employs the GRPO (Group Relative Policy Optimization) algorithm to guide the model in exploring the correct generation path. Extensive experiments show that our framework achieves significant improvements of total +17.0 points on evaluation benchmark, demonstrating the effectiveness of rubric-guided reinforcement learning for open-ended QA.

Out-of-distribution (OOD) detection is critical for the safe deployment of machine learning systems in safety-sensitive domains. Diffusion models have recently emerged as powerful generative models, capable of capturing complex data distributions through iterative denoising. Building on this progress, recent work has explored their potential for OOD detection. We propose *EigenScore*, a new OOD detection method that leverages the eigenvalue spectrum of the posterior covariance induced by a diffusion model. We argue that posterior covariance provides a consistent signal of distribution shift, leading to larger trace and leading eigenvalues on OOD inputs, yielding a clear spectral signature. We further provide analysis explicitly linking posterior covariance to distribution mismatch, establishing it as a reliable signal for OOD detection. To ensure tractability, we adopt a Jacobian-free subspace iteration method to estimate the leading eigenvalues using only forward evaluations of the denoiser. Empirically, EigenScore achieves state-of-the-art performance, with up to 5% AUROC improvement over the best baseline. Notably, it remains robust in near-OOD settings such as CIFAR-10 vs CIFAR-100, where existing diffusion-based methods often fail.

Protein-protein interactions (PPIs) are mediated at the residue level. Most sequence-based PPI models consider residue-residue interactions across two proteins, which can yield accurate interaction scores but are too slow to scale. At proteome scale, identifying candidate PPIs requires evaluating nearly *all possible protein pairs*. For $N$ proteins of average length $L$, exhaustive all-against-all search requires $\mathcal{O}(N^2L^2)$ computation, rendering conventional approaches computationally impractical. We introduce RaftPPI, a scalable framework that approximates residue-level PPI modeling while enabling efficient large-scale retrieval. RaftPPI represents residue interactions with a Gaussian kernel, approximated efficiently via structured random Fourier features, and applies a low-rank factorized attention mechanism that admits pooling into a compact embedding per protein. Each protein is encoded once into an indexable embedding, allowing approximate nearest-neighbor search to replace exhaustive pairwise scoring, reducing proteome-wide retrieval from *months* to *minutes* on a single GPU. On the human proteome with the D-SCRIPT dataset, RaftPPI retrieves the top 20\% candidate pairs ($\sim$200M) in 6 GPU minutes, covering 75.1\% of the true interacting pairs, compared to 4.9 GPU months for the best prior method (61.2\%). Across seven benchmarks with sequence- and degree-controlled splits, RaftPPI achieves state-of-the-art PPI classification and retrieval performance, while enabling residue-aware, retrieval-friendly screening at proteome scale.

While protein language models (PLMs) are one of the most promising avenues of research for future de novo protein design, the way in which they transform sequences to hidden representations, as well as the information encoded in such representations is yet to be fully understood. Several works have attempted to propose interpretability tools for PLMs, but they have focused on understanding how individual sequences are transformed by such models. Therefore, the way in which PLMs transform the whole space of sequences along with their relations is still unknown. In this work we attempt to understand this transformed space of sequences by identifying protein structure and representation with square-root velocity (SRV) representations and graph filtrations. Both approaches naturally lead to a metric space in which pairs of proteins or protein representations can be compared with each other. We analyze different types of proteins from the SCOP dataset and show that the Karcher mean and effective dimension of the SRV shape space follows a non-linear pattern as a function of the layers in ESM2 models of different sizes. Furthermore, we use graph filtrations as a tool to study the context lengths at which models encode the structural features of proteins. We find that PLMs preferentially encode immediate as well as local relations between amino acids, but start to degrade for larger context lengths. The most structurally faithful encoding tends to occur close to, but before the last layer of the models, indicating that training a folding model ontop of these layers might lead to improved folding performance.

CLIP (Contrastive Language-Image Pretraining) has become a popular choice for various downstream tasks. However, recent studies have questioned its ability to represent compositional concepts effectively. These works suggest that CLIP often acts like a bag-of-words (BoW) model, interpreting images and text as sets of individual concepts without grasping the structural relationships. In particular, CLIP struggles to correctly bind attributes to their corresponding objects when multiple objects are present in an image or text. In this work, we investigate why CLIP exhibits this BoW-like behavior. Our key finding is that CLIP does not lack binding information. Through linear probing, robustness tests with increasing object counts, and conjunctive search experiments, we show that attribute–object bindings are already encoded within CLIP’s text and image embeddings. The weakness lies in the cross-modal alignment, which fails to preserve this information. We show it can be accessed cross-modally with a simple linear transformation to text embeddings. This improves CLIP’s attribute-object binding performance and confirms that the information was already encoded unimodally. In practice, this means CLIP-based systems can be enhanced with a lightweight linear layer trained on existing embeddings, avoiding costly encoder retraining.

Identifying how training samples influence/impact Large Language Model (LLM) decision-making is essential for effectively interpreting model decisions and auditing large-scale datasets. Current training sample influence estimation methods (also known as influence functions) undertake this goal by utilizing information flow through the model via its first-order and higher-order gradient terms. However, owing to the large model sizes of today consisting of billions of parameters, these influence computations are often restricted to some subset of model layers to ensure computational feasibility. Prior seminal work by Yeh et al. (2022) in assessing which layers are best suited for computing language data influence concluded that the first (embedding) layers are the most informative for this purpose, using a hypothesis based on influence scores canceling out (i.e., the cancellation effect). In this work, we propose theoretical and empirical evidence demonstrating how the cancellation effect is unreliable, and that middle attention layers are better estimators for influence. Furthermore, we address the broader challenge of aggregating influence scores across layers, and showcase how alternatives to standard averaging (such as ranking and vote-based methods) can lead to significantly improved performance. Finally, we propose better methods for evaluating influence score efficacy in LLMs without undertaking model retraining, and propose a new metric known as the Noise Detection Rate (NDR) that exhibits strong predictive capability compared to the cancellation effect. Through extensive experiments across LLMs of varying types and scales, we concretely determine that the first (layers) are not necessarily better than the last (layers) for LLM influence estimation, contrasting with prior knowledge in the field.

Reasoning-based image quality assessment (IQA) models trained through reinforcement learning (RL) exhibit exceptional generalization, yet the underlying mechanisms and critical factors driving this capability remain underexplored in current research. Moreover, despite their superior performance, these models incur inference energy usage and latency orders of magnitude higher than their earlier counterparts, restricting their deployment in specific scenarios. Through extensive experiments, this paper verifies and elaborates that through RL training, MLLMs leverage their reasoning capability to convert redundant visual representations into compact, cross-domain aligned text representations. This conversion is precisely the source of the generalization exhibited by these reasoning-based IQA models. Building on this fundamental insight, we propose a novel algorithm, RALI, which employs contrastive learning to directly align images with these generalizable text representations learned by RL. This approach eliminates the reliance on reasoning processes and even obviates the need to load an LLM. For the quality scoring task, this framework achieves generalization performance comparable to reasoning-based models while requiring less than 5% of their model parameters and inference time.

High computational costs and slow inference hinder the practical application of video generation models. While prior works accelerate the generation process through feature caching, they often suffer from notable quality degradation. In this work, we reveal that this issue arises from their inability to distinguish truly redundant features, which leads to the unintended skipping of computations on important features. To address this, we propose \textbf{PreciseCache}, a plug-and-play framework that precisely detects and skips truly redundant computations, thereby accelerating inference without sacrificing quality. Specifically, PreciseCache contains two components: LFCache for step-wise caching and BlockCache for block-wise caching. For LFCache, we compute the Low-Frequency Difference (LFD) between the prediction features of the current step and those from the previous cached step. Empirically, we observe that LFD serves as an effective measure of step-wise redundancy, accurately detecting highly redundant steps whose computation can be skipped through reusing cached features. To further accelerate generation within each non-skipped step, we propose BlockCache, which precisely detects and skips redundant computations at the block level within the network. Extensive experiments on various backbones demonstrate the effectiveness of our PreciseCache, which achieves an average of $2.6\times$ speedup without noticeable quality loss. Source code will be released.