ICLR 2026 论文列表

Knowledge Editing has emerged as a promising solution for efficiently updating embedded knowledge in large language models (LLMs). While existing approaches demonstrate effectiveness in integrating new knowledge and preserving the original capabilities of LLMs, they fail to maintain fine-grained irrelevant knowledge facts that share the same subject as edited knowledge but differ in relation and object. This challenge arises because subject representations inherently encode multiple attributes, causing the target and fine-grained irrelevant knowledge to become entangled in the representation space, and thus vulnerable to unintended alterations during editing. To address this, we propose DiKE, a novel approach that Disentangles Knowledge representations for LLM Editing (DiKE). DiKE consists of two key components: a Knowledge Representation Disentanglement (KRD) module that decomposes the subject representation into target-knowledgerelated and -unrelated components, and a Disentanglement-based Knowledge Edit (DKE) module that updates only the target-related component while explicitly preserving the unrelated one. We further derive a closed-form, rank-one parameter update based on matrix theory to enable efficient and minimally invasive edits. To rigorously evaluate fine-grained irrelevant knowledge preservation, we construct FINE-KED, a new benchmark comprising fine-grained irrelevant knowledge at different levels of relational similarity to the edited knowledge. Extensive experiments across multiple LLMs demonstrate that DiKE substantially improves fine-grained irrelevant knowledge preservation while maintaining competitive general editing performance.

Reconstruction techniques, such as 3D Gaussian Splatting (3DGS), are increasingly used for generating scenarios in autonomous driving system (ADS) research. Existing 3DGS-based works for autonomous driving scenario generation have, through various optimizations, achieved high visual similarity in reconstructed scenes. However, this route is built on a strong assumption: that higher scene similarity directly translates into better preservation of ADS behaviour. Unfortunately, this assumption has not been effectively validated, and ADS behaviour is more closely related to the objects within the field of view rather than the global image. Thus, we focus on the perception module—the entry point of ADS. Preliminary experiments reveal that although current methods can produce reconstructions with high overall similarity, they often fail to ensure that the perception module outputs remain consistent with those obtained from the original images. Such a limitation can significantly harm the applicability of reconstruction in the ADS domain. To address this gap, we propose two complementary solutions: a perception-aligned loss, which directly leverages the output differences between reconstructed and ground truth images during the training process; and an object zone quality loss, which specifically reinforces the training on object locations identified by the perception model on ground-truth images. Experiments demonstrate that both of our methods improve the ability of reconstructed scenes to preserve consistency between the perception module outputs and those from the ground-truth inputs.

Inference-time scaling enhances a model’s reasoning by extending its chain-of-thought (CoT). However, existing approaches typically rely on a single policy trained with outcome-reward reinforcement learning (RL), which often suffers from long-horizon plan failures, i.e., the implicit plan drifts away from any valid strategy. This problem is particularly severe for smaller language models (LMs) with long CoTs due to their limited capacity. To address this, we propose Multi-Level Reasoning (MLR), which reformulates long-CoT generation as a two-level stochastic process. Specifically, MLR employs two policies: a high-level planner that generates step descriptors (abstract subgoals) and a low-level executor that produces detailed content conditioned on these descriptors. The planner then generates the next subgoal based on the summarized current step, forming an alternating plan–execute loop. To maintain scalability, we adopt a minimal design, where the base model serves as the low-level policy and a lightweight LoRA module implements the high-level policy. For training, we observe that outcome-reward RL is inefficient and weakly informative for long trajectories (e.g., those exceeding 4K tokens). To overcome this, we introduce online Step-DPO, a process-level preference optimization scheme that leverages Twisted Sequential Monte Carlo (TSMC) to provide scalable stepwise supervision. This yields more effective training, improved stability, and higher accuracy. Extensive experiments on challenging math, science, and logical reasoning benchmarks show that, with only 10% SFT data and 5% of preference data, MLR outperforms both the DeepSeek-R1 distillation and the outcome-reward RL baselines across multiple base models and tasks. More importantly, MLR exhibits slower performance degradation on long-horizon reasoning, demonstrating stronger robustness under extended CoT generation.

Vision-language-action models (VLAs) have garnered significant attention for their potential in advancing robotic manipulation. However, previous approaches predominantly rely on the general comprehension capabilities of vision-language models (VLMs) to generate action signals, often overlooking the rich temporal and causal structure embedded in visual observations. In this paper, we present UniVLA, a unified and native multimodal VLA model that autoregressively models vision, language, and action signals as discrete token sequences. This tokenized formulation naturally supports flexible multimodal task learning, particularly from large-scale video data, and further demonstrates that generative vision supervision can significantly enhance visual understanding. By incorporating world modeling during post-training, UniVLA captures causal dynamics from videos, facilitating effective transfer to downstream policy learning—especially for long-horizon tasks. Our approach sets new state-of-the-art results across several widely used simulation benchmarks, including CALVIN, LIBERO, and Simplenv-Bridge, substantially outperforming prior methods. For example, UniVLA achieves 95.5% average success rate on LIBERO benchmark, surpassing π₀-FAST's 85.5%. We further demonstrate its broad applicability through experiments on real-world ALOHA manipulation tasks and autonomous driving scenarios.

We present WinT3R, a feed-forward reconstruction model capable of online prediction of precise camera poses and high-quality point maps. Previous methods suffer from a trade-off between reconstruction quality and real-time performance. To address this, we first introduce a sliding window mechanism that ensures sufficient information exchange among frames within the window, thereby improving the quality of geometric predictions without introducing a large amount of extra computation. In addition, we leverage a compact representation of cameras and maintain a global camera token pool, which enhances the reliability of camera pose estimation without sacrificing efficiency. These designs enable WinT3R to achieve state-of-the-art performance in terms of online reconstruction quality, camera pose estimation, and reconstruction speed, as validated by extensive experiments on diverse datasets.

It has long been conjectured and empirically observed that neural networks tend to preserve clustering structure. This paper formalizes this conjecture. Specifically, we establish precise conditions for cluster structure preservation and derive bounds to quantify its extent. Through this analysis we are able to show that certain neural networks are learning parameters that preserve the clustering structure of the original data in their embeddings, without the need to impose mechanisms to promote this behavior. Extensive numerical analysis and experiments validate our results. Our findings offer deeper insight into neural network behavior, explaining why certain data types (such as images, audio, and text) benefit more from deep learning. Beyond theory, our findings guide better initialization, feature encoding, and regularization strategies.

Large language models (LLMs) rarely admit uncertainty, often producing fluent but misleading answers, rather than abstaining (i.e., refusing to answer). This weakness is even evident in temporal question answering (QA), where models frequently ignore time-sensitive evidence and conflate facts across different time-periods. In this paper, we present the first empirical study of training LLMs with abstention ability while reasoning about temporal QA. Existing approaches such as calibration might be unreliable in capturing uncertainty in complex reasoning. We instead frame abstention as a teachable skill and introduce a pipeline that couples Chain-of-Thought (CoT) supervision with Reinforcement Learning (RL) guided by abstention-aware rewards. Our goal is to systematically analyze how different information types and training techniques affect temporal reasoning with abstention behavior in LLMs. Through extensive experiments studying various methods, we find that RL yields strong empirical gains on reasoning: a model initialized by Qwen2.5-1.5B-Instruct surpasses GPT-4o by 3.46% and 5.80% in Exact Match on TimeQA-Easy and Hard, respectively. Moreover, it improves the True Positive rate on unanswerable questions by 20% over a pure supervised fine-tuned (SFT) variant. Beyond performance, our analysis shows that SFT induces overconfidence and harms reliability, while RL improves prediction accuracy but exhibits similar risks. Finally, by comparing implicit reasoning cues (e.g., original context, temporal sub-context, knowledge graphs) with explicit CoT supervision, we find that implicit information provides limited benefit for reasoning with abstention. Our study provides new insights into how abstention and reasoning can be jointly optimized, providing a foundation for building more reliable LLMs.

Large language models (LLMs) face growing challenges in efficient generative inference due to the increasing memory demands of Key-Value (KV) caches, especially for long sequences. Existing eviction methods typically retain KV pairs with high attention weights but overlook the impact of attention redistribution caused by token removal, as well as the spatial-temporal dynamics in KV selection. In this paper, we propose ReST-KV, a robust KV eviction method that combines layer-wise output **Re**construction and **S**patial-**T**emporal smoothing to provide a more comprehensive perspective for the KV cache eviction task. Specifically, ReST-KV formulates KV cache eviction as an optimization problem that minimizes output discrepancies through efficient layer-wise reconstruction. By directly modeling how each token’s removal affects the model output, our method naturally captures attention redistribution effects, going beyond simplistic reliance on raw attention weights. To further enhance robustness, we design exponential moving average smoothing to handle temporal variations and an adaptive window-based mechanism to capture spatial patterns. Our method, ReST-KV, significantly advances performance on long-context benchmarks. It surpasses state-of-the-art baselines by 2.58\% on LongBench and 15.2\% on RULER. Additionally, ReST-KV consistently outperforms existing methods on Needle-in-a-Haystack and InfiniteBench, all while achieving a remarkable 10.61$\times$ reduction in decoding latency at 128k context length. The code is included in the supplementary material and is designed for easy reproduction.

The alignment of Large Language Models (LLMs) and brain activity provides a powerful framework to advance our understanding of cognitive neuroscience and artificial intelligence. In this work, we zoom into one of the fundamental units of LLMs—the transformer block—to provide the first systematic computational neuroanatomy of its internal operations and human brain acitivity during language processing. Analyzing 21 state-of-the-art LLMs across five model families, we extract and evaluate 13 distinct intermediate states per transformer block—from initial layer normalization through attention mechanisms to feed-forward networks (FFNs). Our analysis reveals three key findings: (1) The commonly used hidden states in LLMs are surprisingly suboptimal, with over 90\% of brain voxels in sensory and language regions better explained by previously unexplored intermediate computations; (2) Different computational stages within a single transformer block map to anatomically distinct brain systems, revealing an intra-block hierarchy where early attention states align with sensory cortices while later FFN states correspond to association areas—mirroring the cortical processing hierarchy; (3) Rotary Positional Embeddings (RoPE) specifically enhance alignment along the brain's auditory processing streams. Per-head queries with RoPE best explain 74\% of auditory cortex activity compared to 8\% without RoPE, providing the first neurobiological validation of this architectural component in LLMs. Building on these insights, we propose MindTransformer, a feature selection framework that learns brain-aligned representations from all intermediate states. MindTransformer achieves significant brain alignment performance, with correlation improvements in primary auditory cortex exceeding gains from 456× model scaling. Our computational neuroanatomy approach opens new directions for understanding both biological intelligence through the lens of transformer computations and artificial intelligence through principles of brain organization.

As large language models (LLMs) continue to revolutionize AI research, there is a growing interest in building large-scale brain foundation models to advance neuroscience. While most existing brain foundation models are pre-trained on time-series signals or connectome features, we propose a novel graph-based pre-training paradigm for constructing a brain graph foundation model. In this paper, we introduce the Brain Graph Foundation Model, termed BrainGFM, a unified framework that leverages graph contrastive learning and graph masked autoencoders for large-scale fMRI-based pre-training. BrainGFM is pre-trained on a diverse mixture of brain atlases with varying parcellations, significantly expanding the pre-training corpus and enhancing the model’s ability to generalize across heterogeneous fMRI-derived brain representations. To support efficient and versatile downstream transfer, we integrate both graph prompts and language prompts into the model design, enabling BrainGFM to flexibly adapt to a wide range of atlases, neurological and psychiatric disorders, and task settings. Furthermore, we employ meta-learning to optimize the graph prompts, facilitating strong generalization to previously unseen disorders under both few-shot and zero-shot learning conditions via language-guided prompting. BrainGFM is established on 27 neuroimaging datasets spanning 25 common neurological and psychiatric disorders, encompassing 2 types of brain atlases (functional and anatomical) across 8 widely used parcellations, and covering over 25,000 subjects, 60,000 fMRI scans, and a total of 400,000 graph samples aggregated across all atlases and parcellations. The code is available at https://github.com/weixinxu666/BrainGFM.

While Large Language Models have been used to produce interpretable stock forecasts, they mainly focus on analyzing textual reports but not historical price data, also known as Technical Analysis. This task is challenging as it switches between domains: the stock price inputs and outputs lie in the time-series domain, while the reasoning step should be in natural language. In this work, we introduce Verbal Technical Analysis (VTA), a novel framework that combine verbal and latent reasoning to produce stock time-series forecasts that are both accurate and interpretable. To reason over time-series, we convert stock price data into textual annotations and optimize the reasoning trace using an inverse Mean Squared Error (MSE) reward objective. To produce time-series outputs from textual reasoning, we condition the outputs of a time-series backbone model on the reasoning-based attributes. Experiments on stock datasets across U.S., Chinese, and European markets show that VTA achieves state-of-the-art forecasting accuracy, while the reasoning traces also perform well on evaluation metrics judged by industry experts.

Embodied cognition argues that intelligence arises from continuous sensorimotor interaction with the world. It raises an intriguing question: do modern vision-language models (VLMs), trained largely in a disembodied manner, exhibit signs of embodied cognition? To investigate this, we introduce **ENACT**, a benchmark that probes this question through world modeling from egocentric interaction. Grounded in a partially observable Markov decision process (POMDP) framework, **ENACT** comprises two complementary sequence reordering tasks: forward world modeling (predicting an ordered sequence of future states from actions) and inverse world modeling (inferring an ordered sequence of actions from state changes). Correctly solving these tasks indicates that the model has a solid understanding of how the environment will evolve given one's actions. Our scalable dataset contains 8,972 QA pairs derived from diverse, long-horizon household activities in the BEHAVIOR simulator. Experiments reveal a significant performance gap between state-of-the-art VLMs and humans, which widens dramatically as interaction horizons lengthen. We find that models consistently solve the inverse problem better than the forward one and exhibit strong embodied biases, showing a preference for right-handed actions and performance degradation with camera perspectives that deviate from those of human vision. Code and supplementary materials are available in our [anonymous repository](https://github.com/iclrsubmission2026/iclr-2026-submission).

Embodied navigation agents powered by large language models have shown strong performance on individual tasks but struggle to continually acquire new navigation skills, which suffer from catastrophic forgetting. We formalize this challenge as lifelong embodied navigation learning (LENL), where an agent is required to adapt to a sequence of navigation tasks spanning multiple scenes and diverse user instruction styles, while retaining previously learned knowledge. To tackle this problem, we propose Uni-Walker, a lifelong embodied navigation framework that decouples navigation knowledge into task-shared and task-specific components with Decoder Extension LoRA (DE-LoRA). To learn the shared knowledge, we design a knowledge inheritance strategy and an experts co-activation strategy to facilitate shared knowledge transfer and refinement across multiple navigation tasks. To learn the specific knowledge, we propose an expert subspace orthogonality constraint together and a navigation-specific chain-of-thought reasoning mechanism to capture specific knowledge and enhance instruction-style understanding. Extensive experiments demonstrate the superiority of Uni-Walker for building universal embodied navigation agents with lifelong learning. We also provide the code of this work in the Supplementary Materials.

Automated red-teaming has emerged as an essential approach for identifying vulnerabilities in large language models (LLMs). However, most existing methods rely on fixed attack templates and focus primarily on individual high-severity flaws,limiting their adaptability to evolving defenses and their ability to detect complex, high-exploitability vulnerabilities. To address these limitations, we propose AUTO-RT, a reinforcement learning framework designed for automatic jailbreak strategy exploration, i.e., discovering diverse and effective prompts capable of bypassing the safety restrictions of LLMs. AUTO-RT autonomously explores and optimizes attack strategies by interacting with the target model and generating crafted queries that trigger security failures. Specifically, AUTO-RT introduces two key techniques to improve exploration efficiency and attack effectiveness: 1) Dynamic Strategy Pruning, which focuses exploration on high-potential strategies by eliminating highly redundant paths early, and 2) Progressive Reward Tracking, which leverages intermediate downgrade models and a novel First Inverse Rate (FIR) metric to smooth sparse rewards and guide learning. Extensive experiments across diverse white-box and black-box LLM settings demonstrate that AUTO-RT significantly improves success rates (by up to 16.63%), expands vulnerability coverage, and accelerates discovery compared to existing methods.

Modern AI agents such as large language models are trained on diverse tasks---translation, code generation, mathematical reasoning, and text prediction---simultaneously. A key question is to quantify how each individual training task influences performance on a target task, a problem we refer to as task attribution. The direct approach, leave-one-out retraining, measures the effect of removing each task, but is computationally infeasible at scale. An alternative approach that builds surrogate models to predict a target task's performance for any subset of training tasks has emerged in recent literature. Prior work focuses on linear surrogate models, which capture first-order relationships, but miss nonlinear interactions such as synergy, antagonism, or XOR-type effects. In this paper, we first consider a unified task weighting framework for analyzing task attribution methods, and show a new connection between linear surrogate models and influence functions through a second-order analysis. Then, we introduce kernel surrogate models, which more effectively represent second-order task interactions. To efficiently learn the kernel surrogate, we develop a gradient-based estimation procedure that leverages a first-order approximation of pretrained models; empirically, this yields accurate surrogate estimates with less than 2% relative error without repeated retraining. Experiments across multiple domains---including mathematical reasoning in transformers, in-context learning, and multi-objective reinforcement learning---demonstrate the effectiveness of kernel surrogate models. They achieve a 25% higher correlation with the leave-one-out ground truth than linear surrogates and influence-function baselines, enabling more accurate and scalable task attribution. When used for downstream task selection, kernel surrogate models further yield a 40% improvement in demonstration selection for in-context learning and multi-objective reinforcement learning benchmarks.

Reinforcement learning (RL) policies represented in Reproducing Kernel Hilbert Spaces (RKHS) offer powerful representational capabilities. While second-order optimization methods like Newton's method demonstrate faster convergence than first-order approaches, current RKHS-based policy optimization remains constrained to first-order techniques. This limitation stems primarily from the intractability of explicitly computing and inverting the infinite-dimensional Hessian operator in RKHS. We introduce Policy Newton in RKHS, the first second-order optimization framework specifically designed for RL policies represented in RKHS. Our approach circumvents direct computation of the inverse Hessian operator by optimizing a cubic regularized auxiliary objective function. Crucially, we leverage the Representer Theorem to transform this infinite-dimensional optimization into an equivalent, computationally tractable finite-dimensional problem whose dimensionality scales with the trajectory data volume. We establish theoretical guarantees proving convergence to a local optimum with a local quadratic convergence rate. Empirical evaluations on a toy financial asset allocation problem validate these theoretical properties, while experiments on standard RL benchmarks demonstrate that Policy Newton in RKHS achieves superior convergence speed and higher episodic rewards compared to established first-order RKHS approaches and parametric second-order methods. Our work bridges a critical gap between non-parametric policy representations and second-order optimization methods in reinforcement learning.

As fine-tuning becomes impractical at scale, probing is emerging as the preferred evaluation protocol. However, standard linear probing can understate the capability of models whose pre-training optimizes local representations rather than an explicit global representation. This motivates attentive probing, an alternative that uses attention to selectively aggregate patch-level features. Despite growing adoption, attentive probing is still underexplored: existing approaches are often over-parameterized and computationally inefficient. In this work, we revisit attentive probing through the lens of the accuracy vs. parameter-efficiency trade-off. We present the first comprehensive study of existing methods, analyzing their design choices and benchmarking their performance. Building on these insights, we propose efficient probing (EP), a lightweight yet effective multi-query cross-attention mechanism that eliminates redundant projections and reduces the number of trainable parameters. Across multiple benchmarks and pre-training paradigms, EP consistently outperforms linear probing and previous attentive probing methods, and remains effective when combined with parameter-efficient fine-tuning. Beyond evaluation, our analysis uncovers emerging properties of EP, including complementary attention maps, which open new directions for leveraging probing beyond protocol design. Project page: https://vrg.fel.cvut.cz/ep/.

A consistent trend throughout the research of oriented object detection (OOD) has been the pursuit of maintaining comparable performance with fewer and weaker annotations. This is particularly crucial in the remote sensing domain, where the dense object distribution and a wide variety of categories contribute to prohibitively high costs. Based on the supervision level, existing OOD algorithms can be broadly grouped into fully supervised, semi-supervised, and weakly supervised methods. Within the scope of this work, we further categorize them to include sparsely supervised and partially weakly-supervised methods. To address the challenges of large-scale labeling, we introduce the first Sparse Partial Weakly-Supervised Oriented Object Detection (SPWOOD) framework, designed to efficiently leverage only a few sparse weakly-labeled data and plenty of unlabeled data. Our framework incorporates three key innovations: (1) We design a Sparse-annotation-Orientation-and-Scale-aware Student (SOS-Student) model to separate unlabeled objects from the background in a sparsely-labeled setting, and learn orientation and scale information from orientation-agnostic or scale-agnostic weak annotations. (2) We construct a novel Multi-level Pseudo-label Filtering (MPF) strategy that leverages the distribution of model predictions, which is informed by the model’s multi-layer predictions. (3) We propose a unique sparse partitioning approach, ensuring equal treatment for each category. Extensive experiments on the DOTA-v1.0 and v1.5 datasets show that SPWOOD framework achieves a significant performance gain over traditional OOD methods mentioned above, offering a highly cost-effective solution. Our code will be public soon.

Batch size scheduling (BSS) plays a critical role in large-scale deep learning training, influencing both optimization dynamics and computational efficiency. Yet, its theoretical foundations remain poorly understood. In this work, we show that the **functional scaling law (FSL)** framework introduced in [Li et al. (2025a)](https://arxiv.org/abs/2509.19189) provides a principled lens for analyzing BSS. Specifically, we characterize the optimal BSS under a fixed data budget and show that its structure depends sharply on task difficulty. For easy tasks, optimal schedules keep increasing batch size throughout. In contrast, for hard tasks, the optimal schedule maintains small batch sizes for most of training and switches to large batches only in a late stage. To explain the emergence of late switching, we uncover a dynamical mechanism—the **fast catch-up** effect—which also manifests in large language model (LLM) pretraining. After switching from small to large batches, the loss rapidly aligns with the constant large-batch trajectory. Using FSL, we show that this effect stems from rapid forgetting of accumulated gradient noise, with the catch-up speed determined by task difficulty. Crucially, this effect implies that *large batches can be safely deferred to late training* without sacrificing performance, while substantially reducing data consumption. Finally, extensive LLM pretraining experiments—covering both Dense and MoE architectures with up to **1.1B** parameters and **1T** tokens—validate our theoretical predictions. Across all settings, late-switch schedules consistently outperform constant-batch and early-switch baselines.