ICLR 2026 Papers

Transformer architectures have achieved remarkable success in various domains. While efficient alternatives to Softmax Attention have been widely studied, the search for more expressive mechanisms grounded in theoretical insight—even at greater computational cost—has been relatively underexplored. In this work, we bridge this gap by proposing Local Linear Attention (LLA), a novel attention mechanism derived from nonparametric statistics through the lens of test-time regression. First, we show that LLA offers theoretical advantages over Linear and Softmax Attention for associative memory via a bias-variance trade-off analysis. Next, we address its computational challenges and propose two memory-efficient primitives to tackle the $\Theta(n^2d)$ and $\Theta(nd^2)$ complexity. We then introduce {FlashLLA}, a hardware-efficient, blockwise algorithm that enables scalable and parallel computation on modern accelerators. In addition, we implement and profile a customized inference kernel that significantly reduces memory overheads. Finally, we empirically validate the advantages and limitations of LLA on test-time regression, in-context regression, associative recall and state tracking tasks. Experiment results demonstrate that LLA effectively adapts to non-stationarity, outperforming strong baselines in test-time training and in-context learning, and exhibiting promising evidence for its scalability and applicability in large-scale models.

Recent advances in inference-time scaling, particularly those leveraging reinforcement learning with verifiable rewards, have substantially enhanced the reasoning capabilities of Large Vision-Language Models (LVLMs). Inspired by this success, similar strategies have been applied to multimodal reasoning, yet their impact on visual perception remains unclear. To investigate this gap, we introduce DisTANCE, a perception-centric benchmark for visual estimation tasks. Evaluation results show that LVLMs exhibit limited estimation precision, and inference-time scaling offers only marginal gains. We attribute this to the fast perception paradigm of current LVLMs, where visual understanding is treated as a one-shot output without modeling the underlying perceptual process. To address this, we propose Perception-Time Scaling (PTS), a novel paradigm that encourages token-rich perception and decomposes complex perception problems into intermediate tractable sub-problems, thereby enabling perception to align with and benefit from inference-time scaling. Combined with reinforcement learning techniques, PTS significantly improves perception accuracy, raising high-precision performance on DisTANCE from 8.0% to 64.7%, and generalizes well to out-of-domain tasks. Surprisingly, even though PTS data are purely synthetic, combining them with math reasoning data yields consistent gains in both reasoning and real-world perception benchmarks. Further analysis reveals that PTS introduces more perception-related tokens and increases the model’s attention to image tokens. Our code and data will be publicly released.

Large language models remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Defending against novel jailbreaks represents a critical challenge in AI safety. Adversarial training---designed to make models robust against worst-case perturbations---has been the dominant paradigm for adversarial robustness. However, due to optimization challenges and difficulties in defining realistic threat models, adversarial training methods often fail on newly developed jailbreaks in practice. This paper proposes a new paradigm for improving robustness against unseen jailbreaks, centered on the Adversarial Déjà Vu hypothesis: novel jailbreaks are not fundamentally new, but largely recombinations of adversarial skills from previous attacks. We study this hypothesis through a large-scale analysis of 32 attack papers published over two years. Using an automated pipeline, we extract and compress adversarial skills into a sparse dictionary of primitives, with LLMs generating human-readable descriptions. Our analysis reveals that unseen attacks can be effectively explained as sparse compositions of earlier skills, with explanatory power increasing monotonically as skill coverage grows. Guided by this insight, we introduce Adversarial Skill Compositional Training (ASCoT), which trains on diverse compositions of skill primitives rather than isolated attack instances. ASCoT substantially improves robustness to unseen attacks, including multi-turn jailbreaks, while maintaining low over-refusal rates. We also demonstrate that expanding adversarial skill coverage, not just data scale, is key to defending against novel attacks.

The process through which humans perceive and learn visual representations in dynamic environments is highly complex. From a structural perspective, the human eye decouples the functions of cone and rod cells: cones are primarily responsible for color perception, while rods are specialized in detecting motion, particularly variations in light intensity. These two distinct modalities of visual information are integrated and processed within the visual cortex, thereby enhancing the robustness of the human visual system. Inspired by this biological mechanism, modern hardware systems have evolved to include not only color-sensitive RGB cameras but also motion-sensitive Dynamic Visual Systems, such as spike cameras. Building upon these advancements, this study seeks to emulate the human visual system by integrating decomposed multi-modal visual inputs with modern latent-space generative frameworks. We named it ***SpikeGen***. We evaluate its performance across various spike-RGB tasks, including conditional image and video deblurring, dense frame reconstruction from spike streams, and high-speed scene novel-view synthesis. Supported by extensive experiments, we demonstrate that leveraging the latent space manipulation capabilities of generative models enables an effective synergistic enhancement of different visual modalities, addressing spatial sparsity in spike inputs and temporal sparsity in RGB inputs.

Engineering design operates through hierarchical abstraction from system specifications to component implementations, requiring visual understanding coupled with mathematical reasoning at each level. While Multi-modal Large Language Models (MLLMs) excel at natural image tasks, their ability to extract mathematical models from technical diagrams remains unexplored. We present \textbf{CircuitSense}, a comprehensive benchmark evaluating circuit understanding across this hierarchy through 8,006+ problems spanning component-level schematics to system-level block diagrams. Our benchmark uniquely examines the complete engineering workflow: Perception, Analysis, and Design, with a particular emphasis on the critical but underexplored capability of deriving symbolic equations from visual inputs. We introduce a hierarchical synthetic generation pipeline consisting of a grid-based schematic generator and a block diagram generator with auto-derived symbolic equation labels. Comprehensive evaluation of eight state-of-the-art MLLMs, including both closed-source and open-source models, reveals fundamental limitations in visual-to-mathematical reasoning. Closed-source models achieve over 85\% accuracy on perception tasks involving component recognition and topology identification, yet their performance on symbolic derivation and analytical reasoning falls below 19\%, exposing a critical gap between visual parsing and symbolic reasoning. Models with stronger symbolic reasoning capabilities consistently achieve higher design task accuracy, confirming the fundamental role of mathematical understanding in circuit synthesis and establishing symbolic reasoning as the key metric for engineering competence. Our synthetic pipeline code is available at \href{https://anonymous.4open.science/r/CircuitSense-8AC7/README.md}{URL}.

Large Language Models (LLMs) have seen remarkable advancements, achieving state-of-the-art results in diverse applications. Fine-tuning, an important step for adapting LLMs to specific downstream tasks, typically involves further training on corresponding datasets. However, a fundamental discrepancy exists between current fine-tuning datasets and the token-level optimization mechanism of LLMs: most datasets are designed at the sentence-level, which introduces token-level noise, causing negative influence to final performance. In this paper, we propose XTF, an explainable token-level noise filtering framework. XTF decomposes the complex and subtle contributions of token-level data to the fine-tuning process into three distinct and explicit attributes (reasoning importance, knowledge novelty, and task relevance), which can be assessed using scoring methods, and then masks the gradients of selected noisy tokens accordingly to optimize the performance of fine-tuned LLMs. We conduct extensive experiments on three representative downstream tasks (math, code and medicine) across 7 mainstream LLMs. The results demonstrate that XTF can significantly improve downstream performance by up to 13.7% compared to regular fine-tuning. Our work highlights the importance of token-level dataset optimization, and demonstrates the potential of strategies based on attribute decomposition for explaining complex training mechanisms.

Reinforcement Learning with Verifiable Rewards (RLVR) strengthens LLM reasoning but training often oscillates between {entropy collapse} and {entropy explosion}. We trace both hazards to the mean-baseline used in value-free RL (\eg GRPO/DAPO), which improperly penalizes negative-advantage samples under reward outliers. We propose {Quantile Advantage Estimation} (QAE), replacing the mean with a group-wise $K$-quantile baseline. QAE induces a response-level, two-regime gate: on hard queries ($p \le 1{-}K$) it reinforces rare successes, while on easy queries ($p > 1{-}K$) it targets remaining failures. Under first-order softmax updates, we prove {two-sided entropy safety}, giving lower/upper bounds on one-step entropy change that curb explosion and prevent collapse. Empirically, this minimal modification stabilizes entropy, sparsifies credit assignment (with tuned $K$, roughly 80\% of responses receive zero advantage), and yields sustained pass@1 gains on Qwen3-8B/14B-Base across AIME'24/'25 and AMC'23. These results identify {baseline design}—rather than token-level heuristics—as the primary mechanism for scaling RLVR.

Vision–Language–Action (VLA) agents follow instructions to perform multi-step tasks in multimodal environments. To support planning and execution in such settings, many approaches typically adopt structured post-hoc or rely on fixed decomposition and rigid alignment to improve success rate. However, once an intermediate subgoal or action is mis-specified and without a flexible correction mechanism, local errors propagate through subsequent steps and eventually accumulate into cascading failures in long-horizon reasoning. To mitigate this compounding effect, we propose Reflective Contrastive Alignment and Planning Architecture (ReCAPA), a framework that uses predictive correction to anticipate deviations and adjust representations across three levels: actions, subgoals, and trajectories. Semantic alignment is enforced at all levels using a Sinkhorn-based module and a Score-field module. The corrective signals, derived from predictive correction and alignment mechanisms, jointly update the execution network during training, enabling it to flexibly adjust fine-grained steps to remain aligned with the overall intent. We further introduce two new metrics to quantify error propagation and recovery processes in tasks. Experiments show that ReCAPA achieves competitive results on embodied agent benchmarks such as VisualAgentBench, MineDojo, and MAP-THOR, outperforming strong proprietary and open-source Large Language Model (LLM) baselines.

Contemporary coding-agent benchmarks applaud “first correct answer,” silently assuming infinite tokens, container minutes, and developer patience. In production, every LLM call, test re-run, and rollback incurs hard cost; agents that cannot budget these resources are dead on arrival. We close the gap with USACOArena, an ICPC-inspired arena where agents pay deterministic credits for every prompt, compilation, test, or rollback. A task becomes a cost–benefit negotiation under uncertainty: is a second sample worth 15\% of the remaining budget, or should the agent pivot to a cheaper heuristic? Real-time deduction exposes decision profiles hidden from static leaderboards: the tax of over-specialized generators, the ROI of early-exit heuristics, and the compound interest of lightweight scaffolding. Even identically seeded agents diverge in self-play, revealing a rich policy space where the same model oscillates between spendthrift submission sprees and parsimonious exploration. Released as a reproducible benchmark and zero-shot curriculum, USACOArena provides the traces, credit engine, and six state-of-the-art decision logs to catalyze research on coding agents that know when to stop.

Long document summarization remains a significant challenge for current large language models (LLMs), as existing approaches commonly struggle with information loss, factual inconsistencies, and coherence issues when processing excessively long documents. We propose SummQ, a novel adversarial multi-agent framework that addresses these limitations through collaborative intelligence between specialized agents operating in two complementary domains: summarization and quizzing. Our approach employs summary generators and reviewers that work collaboratively to create and evaluate comprehensive summaries, while quiz generators and reviewers create comprehension questions that serve as continuous quality checks for the summarization process. This adversarial dynamic, enhanced by an examinee agent that validates whether the generated summary contains the information needed to answer the quiz questions, enables iterative refinement through multifaceted feedback mechanisms. We evaluate SummQ on three widely used long document summarization benchmarks. Experimental results demonstrate that our framework significantly outperforms existing state-of-the-art methods across ROUGE and BERTScore metrics, as well as in LLM-as-a-Judge and human evaluations. Our comprehensive analyses reveal the effectiveness of the multi-agent collaboration dynamics, the influence of different agent configurations, and the impact of the quizzing mechanism. This work establishes a new approach for long document summarization that uses adversarial agentic collaboration to improve summarization quality.

Heuristic functions are central to the performance of search algorithms such as A*, where \emph{admissibility}—the property of never overestimating the true shortest-path cost—guarantees solution optimality. Recent deep learning approaches often disregard full admissibility and provide limited guarantees on generalization beyond the training data. We address both of these limitations. First, we pose heuristic learning as a constrained optimization problem and introduce \emph{Cross-Entropy Admissibility (CEA)}, a loss function that enforces admissibility during training. When evaluated on the Rubik’s Cube domain, our method yields heuristics with near-perfect admissibility and significantly stronger guidance than compressed pattern database (PDB) heuristics. On the theoretical side, we derive a new upper bound on the expected suboptimality of A*. By leveraging PDB abstractions and the structural properties of graphs such as the Rubik’s Cube, we tighten the bound on the number of training samples needed for A* to generalize to unseen states. Replacing a general hypothesis class with a ReLU neural network gives bounds that depend primarily on the network’s width and depth, rather than on graph size. Using the same network, we also provide the first generalization guarantees for \emph{goal-dependent} heuristics.

Mean-field games (MFGs) study the Nash equilibrium of systems with a continuum of interacting agents, which can be formulated as the fixed-point of optimal control problems. They provide a unified framework for a variety of applications, including optimal transport (OT) and generative models. Despite their broad applicability, solving high-dimensional MFGs remains a significant challenge due to fundamental computational and analytical obstacles. In this work, we propose a particle-based deep Flow-Matching (FM) method to tackle high-dimensional deterministic MFG computation. In each iteration of our proximal best response scheme, particles are updated using first-order information, and a flow neural network is trained to match the velocity of the sample trajectories in a simulation-free manner. Theoretically, in the optimal control setting, we prove that our scheme converges to a stationary point sublinearly, and upgrade to linear (exponential) convergence under additional convexity assumptions. Our proof uses FM to induce an Eulerian coordinate (density-based) from a Lagrangian one (particle-based), and this also leads to certain equivalence results between the two formulations for MFGs when the Eulerian solution is sufficiently regular. Our method demonstrates promising performance on non-potential MFGs and high-dimensional OT problems cast as MFGs through a relaxed terminal-cost formulation.

Exploration of dynamic environments with sparse rewards is a significant challenge in Reinforcement Learning, often leading to inefficient exploration and brittle policies. To address this, we introduce a novel graph-based intrinsic reward using Effective Resistance, a metric from spectral graph theory. This reward formulation guides the agent to seek configurations that are directly correlated to successful goal reaching states. We provide theoretical guarantees, proving that our method not only learns a robust policy but also achieves faster convergence by serving as a variance reduction baseline to the standard discounted reward formulation. We perform extensive empirical analysis across several challenging environments to demonstrate that our approach significantly outperforms state-of-the-art baselines, demonstrating improvements of up to 59% in success rate, 56% in timesteps taken to reach the goal, and 4 times more accumulated reward. We augment all of the supporting lemmas and theoretically motivated hyperparameter choices with corresponding experiments.

Rotary Positional Embedding (RoPE) is a widely adopted technique for encoding position in language models, which, while effective, causes performance breakdown when input length exceeds training length. Prior analyses assert (rightly) that long inputs cause channels to rotate "out of distribution," but it is not clear how extra rotation relates to or causes pathological behavior. Through empirical analysis we advance a unified geometric understanding of attention behavior with RoPE. We find that attention induces tight clustering of separated key and query latent point clouds, allowing for creation of sink tokens: placeholders that allow attention heads to avoid token mixing when not required. RoPE applied to longer inputs damages this key/query cluster separation, producing pathological behavior by inhibiting sink token functionality. From this geometric perspective, we propose RoPE-ID (In Distribution), a straightforward modification that allows attention layers to generalize to longer inputs out of the box: apply RoPE with high frequency to a subset of channels. We demonstrate the effectiveness of RoPE-ID for extended inputs using 1B and 3B parameter Transformers on the LongBench and RULER information retrieval benchmarks.

Concerns have been raised that knowledge distillation may transfer test-set knowledge from a contaminated teacher to a clean student—a “data laundering” effect that potentially threatens evaluation integrity. In this paper, we assess the severity of this phenomenon. If these concerns regarding data laundering are minor, then distillation could be used to mitigate risks of direct data exposure. Across eight classification benchmarks, we find that substantial laundering is the exception rather than the rule: unlike the large performance gains from direct contamination, any accuracy inflation from laundering is consistently smaller and statistically insignificant in all but two cases. More broadly, we find that the two phenomena are weakly correlated, suggesting that laundering is not simply a diluted form of contamination but a distinct effect that arises primarily when benchmarks exhibit large train–test distribution gaps. Motivated by this, we conduct controlled experiments that systematically enlarge the train–test distance on two benchmarks where laundering was initially negligible, and observe that laundering becomes more significant as the gap widens. Taken together, our results indicate that knowledge distillation, despite rare benchmark-specific residues, can be expected to function as an effective decontamination technique that largely mitigates test-data leakage.

Transformer-based large language models (LLMs) have achieved remarkable success, yet their standard attention mechanism incurs quadratic computation and memory costs with respect to sequence length, posing a major bottleneck for long-context training. Prior work tackles this challenge along two directions: (1) kernel-level optimizations, which accelerate dense and sparse attention operators; and (2) module-level strategies, often referred to as distributed attention or context parallel training, which scale attention across multiple devices. However, systematic evaluation still remains limited: operator-level comparisons are often incomplete, while context parallel strategies are typically framework-specific, with unclear performance analysis across contexts. To address these gaps, we propose a unified benchmark that integrates representative attention kernels and context parallel mechanisms with a modular and extensible interface for evaluation. The benchmark evaluates methods along two critical dimensions: (1) attention mask patterns, which strongly affect efficiency, scalability, and usability, and (2) sequence length and distributed scale, which determine performance under extreme long-context training. Through comprehensive experiments on the cluster of up to 96 GPUs, our benchmark enables reproducible comparisons, highlights method-specific trade-offs, and provides practical guidance for designing and deploying attention mechanisms in long-context LLM training.

Hands play a central role in daily life, yet modeling natural hand motions remains underexplored. Existing methods that tackle text-to-hand-motion generation or hand animation captioning rely on studio-captured datasets with limited actions and contexts, making them costly to scale to “in-the-wild” settings. Further, contemporary models and their training schemes struggle to capture animation fidelity with text–motion alignment. To address this, we (1) introduce ‘3D Hands in the Wild’ (3D-HIW), a dataset of 32K 3D hand-motion sequences and aligned text, and (2) propose CLUTCH, an LLM-based hand animation system with two critical innovations: (a) SHIFT, a novel VQ-VAE architecture to tokenize hand motion, and (b) a geometric refinement stage to finetune the LLM. To build 3D- HIW, we propose a data annotation pipeline that combines vision–language models (VLMs) and state-of-the-art 3D hand trackers, and apply it to a large corpus of egocentric action videos covering a wide range of scenarios. To fully capture motion in-the-wild, CLUTCH employs SHIFT, a part–modality decomposed VQ- VAE, which improves generalization and reconstruction fidelity. Finally, to improve animation quality, we introduce a geometric refinement stage, where CLUTCH is co-supervised with a reconstruction loss applied directly to decoded hand motion parameters. Experiments demonstrate state-of-the-art performance on text-to- motion and motion-to-text tasks, establishing the first benchmark for scalable in-the-wild hand motion modelling. Code, data and models will be released.

Facts change over time, making it essential for Large Language Models (LLMs) to handle time-sensitive factual knowledge accurately and reliably. Although factual Time-Sensitive Question-Answering (TSQA) tasks have been widely developed, existing benchmarks often face manual bottlenecks that limit scalable and comprehensive TSQA evaluation. To address this issue, we propose TDBench, a new benchmark that systematically constructs TSQA pairs by harnessing temporal databases and database techniques, such as temporal functional dependencies, temporal SQL, and temporal joins. We also introduce a new evaluation metric called time accuracy, which assesses the validity of time references in model explanations alongside traditional answer accuracy for a more fine-grained TSQA evaluation. Extensive experiments on contemporary LLMs show how TDBench enables scalable and comprehensive TSQA evaluation while reducing the reliance on human labor, complementing current TSQA evaluation approaches that largely center on Wikipedia/Wikidata by enabling LLM evaluation on application-specific data.

Model merging has emerged as an efficient and flexible paradigm for multi-task learning, with numerous methods being proposed in recent years. However, these state-of-the-art techniques are typically evaluated on benchmark suites that are highly favorable to model merging, and their robustness in more realistic settings remains largely unexplored. In this work, we first investigate the vulnerabilities of model-merging methods and pinpoint the source-model characteristics that critically underlie them. Specifically, we identify two factors that are particularly harmful to the merging process: (1) disparities in task vector norms, and (2) the low confidence of the source models. To address this issue, we propose **DisTaC** (**Dis**tillation for **Ta**sk vector **C**onditioning), a novel method that pre-conditions these problematic task vectors before the merge. DisTaC leverages knowledge distillation to adjust a task vector's norm and increase source-model confidence while preserving its essential task-specific knowledge. Our extensive experiments demonstrate that by pre-conditioning task vectors with DisTaC, state-of-the-art merging techniques can successfully integrate models that exhibit these harmful traits, where they would otherwise fail, and achieve significant performance gains.