ICLR 2026 Papers

Large Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, explore, and refine what they need through multi-turn conversational exchange. Although analysis of LLM conversation logs has confirmed that underspecification occurs frequently in user instructions, LLM evaluation has predominantly focused on the single-turn, fully-specified instruction setting. In this work, we perform large-scale simulation experiments to compare LLM performance in single- and multi-turn settings. Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover.

Model editing techniques are essential for efficiently updating knowledge in large language models (LLMs). However, the effectiveness of existing approaches degrades in massive editing scenarios, particularly when evaluated with practical metrics. Their robustness is also limited in context-rich settings or when editing multiple facts of the same subject simultaneously. We attribute these failures to the embedding misalignment among knowledge items, which undermines editing reliability at scale. To address this, we propose EAMET (Embedding Alignment Model Editing in Transformers), which addresses this issue by aligning the space of key and residual embeddings. Extensive experiments across six LLMs and three datasets demonstrate that EAMET consistently outperforms existing methods, achieving about 90\% editing efficacy when editing 10k facts.

Mathematical reasoning in real-world video presents a fundamentally different challenge than static images or text. It requires interpreting fine-grained visual information, accurately reading handwritten or digital text, and integrating spoken cues, often dispersed non-linearly over time. In such multimodal contexts, success hinges not just on perception, but on selectively identifying and integrating the right details from a rich and noisy stream of content. To this end, we introduce VideoMathQA, a benchmark designed to evaluate whether models can perform such temporally extended cross-modal reasoning on videos. The benchmark spans 10 diverse mathematical domains, covering videos from 10 seconds to over 1 hour. We employ graduate-level experts to ensure high quality, for over 920 man-hours of annotation. To reflect real-world scenarios, questions are designed around three core reasoning challenges: direct problem solving, conceptual transfer, which requires applying learned methods to new problems; and deep instructional comprehension, involving multi-step reasoning over extended explanations and partially worked-out solutions. Each question includes multi-step reasoning annotations, enabling fine-grained diagnosis of model capabilities. Through this benchmark, we establish an evaluation framework for models that must reason, rather than merely perceive, jointly ground concepts across visual, audio, and textual modalities, across temporally extended mathematical problem settings.

While Large Language Models (LLMs) have demonstrated impressive capabilities, their output quality remains inconsistent across various application scenarios, making it difficult to identify trustworthy responses, especially in complex tasks requiring multi-step reasoning. In this paper, we propose a **Tok**en-level **U**ncertainty estimation framework for **R**easoning (**TokUR**) that enables LLMs to self-assess and self-improve their responses in mathematical reasoning. Specifically, we introduce low-rank random weight perturbation during LLM decoding to generate predictive distributions for token-level uncertainty estimation, and we aggregate these uncertainty quantities to capture the semantic uncertainty of generated responses. Experiments on mathematical reasoning datasets of varying difficulty demonstrate that TokUR exhibits a strong correlation with answer correctness and model robustness, and the uncertainty signals produced by TokUR can be leveraged to enhance the model’s reasoning performance at test time. These results highlight the effectiveness of TokUR as a principled and scalable approach for improving the reliability and interpretability of LLMs in challenging reasoning tasks.

Existing multimodal browsing benchmarks often fail to require genuine multimodal reasoning, as many tasks can be solved with text-only heuristics without vision-in-the-loop verification. We introduce MMSearch-Plus, a 311-task benchmark that enforces multimodal understanding by requiring extraction and propagation of fine-grained visual cues through iterative image–text retrieval and cross-validation under retrieval noise. Our curation procedure seeds questions whose answers require extrapolating from spatial cues and temporal traces to out-of-image facts such as events, dates, and venues. Beyond the dataset, we provide a model-agnostic agent framework with standard browsing tools and a set-of-mark (SoM) module, which lets the agent place marks, crop subregions, and launch targeted image/text searches. SoM enables provenance-aware zoom-and-retrieve and improves robustness in multi-step reasoning. We evaluated closed- and open-source MLLMs in this framework. The strongest system achieves an end-to-end accuracy of 36.0%, and integrating SoM produces consistent gains in multiple settings, with improvements up to +3.9 points. From failure analysis, we observe recurring errors in locating relevant webpages and distinguishing between visually similar events. These results underscore the challenges of real-world multimodal search and establish MMSearch-Plus as a rigorous benchmark for advancing agentic MLLMs.

Real-time image segmentation is a key enabler for AR/VR, robotics, drones, and autonomous systems, where tight accuracy, latency, and energy budgets must be met on resource‑constrained edge devices. While U‑Net offers a favorable balance of accuracy and efficiency compared to large transformer‑based models, achieving real‑time performance on high‑resolution input remains challenging due to compute, memory, and power limits. Extreme quantization, particularly binary networks, is appealing for its hardware‑friendly operations. However, two obstacles limit practicality: (1) severe accuracy degradation, and (2) a lack of end‑to‑end implementations that deliver efficiency on general‑purpose GPUs. We make two empirical observations. (1) An explicit zero state is essential: training with zero masking to binary U‑Net weights yields noticeable sparsity. (2) Quantization sensitivity is relatively uniform across layers. Motivated by these findings, we introduce Masked Binary U‑Net (MBU‑Net), obtained through a cost‑aware masking strategy that prioritizes masking where it yields the highest accuracy‑per‑cost, reconciling accuracy with near‑binary efficiency. To realize these gains in practice, we develop a GPU execution framework that maps MBU‑Net to Tensor Cores via a subtractive bit‑encoding scheme, efficiently implementing masked binary weights with binary activations. This design leverages native binary Tensor Core BMMA instructions, enabling high throughput and energy savings on widely available GPUs. Across 3 segmentation benchmarks, MBU‑Net attains near full‑precision accuracy (3\% average drop) while delivering 2.04$\times$ speedup and 3.54$\times$ energy reductions over a 16-bit floating point U‑Net. The code is available at https://github.com/ChunshuWu/MBU-Net.

As large language models (LLMs) are pretrained on massive web corpora, careful selection of data becomes essential to ensure effective and efficient learning. While perplexity (PPL)-based filtering has demonstrated strong performance, it suffers from drawbacks: substantial time costs and inherent unreliability of the model when handling noisy or out-of-distribution samples. In this work, we propose a simple yet powerful alternative: a prior-based data filtering method that estimates token priors using corpus-level term frequency statistics, inspired by linguistic insights on word roles and lexical density. Our approach filters documents based on the mean and standard deviation of token priors, serving as a fast proxy to PPL while requiring no model inference. Despite its simplicity, the prior-based filter achieves the highest average performance across 20 downstream benchmarks, while reducing time cost by over 1000× compared to PPL-based filtering. We further demonstrate its applicability to symbolic languages such as code and math, and its dynamic adaptability to multilingual corpora without supervision.

Video Class-Incremental Learning (VCIL) seeks to develop models that continuously learn new action categories over time without forgetting previously acquired knowledge. Unlike traditional Class-Incremental Learning (CIL), VCIL introduces the added complexity of spatiotemporal structures, making it particularly challenging to mitigate catastrophic forgetting while effectively capturing both frame-shared semantics and temporal dynamics. Existing approaches either rely on exemplar rehearsal, raising concerns over memory and privacy, or adapt static image-based methods that neglect temporal modeling. To address these limitations, we propose Spatiotemporal Preservation and Routing (StPR), a unified and exemplar-free VCIL framework that explicitly disentangles and preserves spatiotemporal information. We begin by introducing Frame-Shared Semantics Distillation (FSSD), which identifies semantically stable and meaningful channels by jointly considering channel-wise sensitivity and classification contribution. By selectively regularizing these important semantic channels, FSSD preserves prior knowledge while allowing for adaptation. Building on this preserved semantic space, we further design a Temporal Decomposition-based Mixture-of-Experts (TD-MoE), which dynamically routes task-specific experts according to temporal dynamics, thereby enabling inference without task IDs or stored exemplars. Through the synergy of FSSD and TD-MoE, StPR progressively leverages spatial semantics and temporal dynamics, culminating in a unified, exemplar-free VCIL framework. Extensive experiments on UCF101, HMDB51, SSv2 and Kinetics400 show that our method outperforms existing baselines while offering improved interpretability and efficiency in VCIL. Code is available in the supplementary materials.

An important aspect subtending language understanding and production is the ability to independently encode positional and symbolic information of the words within a sentence. In Transformers, positional information is typically encoded using Positional Encodings (PEs). One such popular PE, namely Rotary PE (RoPE), has been widely used due to its empirical success. Recently, it has been argued that part of RoPE's success emerges from its ability to encode robust positional and semantic information using large and small frequencies, respectively. In this work, we perform a deeper dive into the positional versus symbolic dichotomy of attention heads behavior, both at the theoretical and empirical level. We provide general definitions of what it means for a head to behave positionally or symbolically, prove that these are two mutually exclusive behaviors and develop a metric to quantify them. We apply our framework to analyze Transformer-based LLMs using RoPE and find that all heads exhibit a strong correspondence between behavior and frequency use. Finally, we introduce canonical tasks designed to be either purely positional or symbolic, and demonstrate that the Transformer performance causally relates to the ability of attention heads to leverage the appropriate frequencies. In particular, we show that we can control the Transformer performance by controlling which frequencies the attention heads can access. Altogether, our work provides a detailed understanding of RoPE, and how its properties relate to model behavior.

Vision–Language Models (VLMs) show promise as zero-shot goal-conditioned value functions, but their frozen pre-trained representations limit generalization and temporal reasoning. We introduce VITA, a zero-shot value function learning method that enhances both capabilities via test-time adaptation. At inference, a lightweight adaptation module is updated via a gradient step on a meta-learned self-supervised loss, such that each test-time update improves value estimation. By updating sequentially over a trajectory, VITA encodes history into its parameters, addressing the temporal reasoning limitations. To mitigate shortcut learning, we propose a dissimilarity-based sampling strategy that selects semantically diverse segments of the trajectory during training. In real-world robotic manipulation tasks, VITA generalizes from a single training environment to diverse out-of-distribution tasks, environments, and embodiments, outperforming the state-of-the-art zero-shot method using autoregressive VLMs. Furthermore, we demonstrate that VITA’s zero-shot value estimates can be utilized for reward shaping in offline reinforcement learning, resulting in multi-task policies on the Meta-World benchmark that exceed the performance of those trained with the simulation’s fuzzy-logic dense rewards.

Vision-language-action (VLA) models aim to understand natural language instructions and visual observations and execute corresponding actions as an embodied agent. Recent advancements have integrated future images into the understanding-action loop, enabling foresight-driven policies that reduce abstract action prediction to a more tractable inverse kinematics problem. However, existing models either rely on external experts for modality unification or treat image generation and action prediction as separate processes, limiting the benefits of direct synergy between these tasks. In this work, we propose Unified Diffusion VLAs, which tightly couple understanding, generation, and action in a mutually reinforcing manner. Our method optimizes the generation of actions and images jointly through a synchronous denoising diffusion process, where action tokens progressively attend to future image tokens. This iterative refinement enables actions to evolve from initialization with sufficient visual guidance, ensuring precise action execution. We introduce a hybrid attention mechanism and the Joint Discrete Denoising Diffusion Process (JD3P), which integrates multiple modalities into a unified trajectory. We also propose a two-stage training pipeline and several inference-time techniques that optimize performance and efficiency. Our approach achieves state-of-the-art performance on benchmarks such as CALVIN, LIBERO, and SimplerEnv, and we demonstrate its effectiveness through ablation studies and real-world evaluations.

Reliable cognitive load (CL) prediction in real-world settings is fundamentally constrained by the poor and variable quality of physiological signals. In safety-critical tasks such as driving, degraded signal quality can severely compromise prediction accuracy, limiting the deployment of existing models outside controlled lab conditions. To address this challenge, we propose CogMoE, a signal quality–guided Mixture-of-Experts (MoE) framework that dynamically adapts to heterogeneous and noisy inputs. CogMoE flexibly integrates physiological modalities, including EEG, ECG, EDA, and gaze, through quality-aware gating, enabling context-sensitive fusion. The framework operates in two stages: (1) quality-aware multi-modal synchronization and recovery to mitigate artifacts, temporal misalignment, and missing data, and (2) signal-quality-specific expert modeling via a cross-modal MoE transformer that regulates information flow based on signal reliability. To further improve stability, we introduce CORTEX Loss, which balances reconstruction fidelity and expert utilization under noise. Experiments on CL-Drive and ADABase show that CogMoE outperforms strong baselines, delivering consistent improvements across diverse signal qualities.

Humans ponder before articulating complex sentence elements, enabling deeper cognitive processing through focused effort. In this work, we introduce this pondering process into language models by repeatedly invoking the forward process within a single token generation step. During pondering, instead of generating an actual token sampled from the prediction distribution, the model ponders by yielding a weighted sum of all token embeddings according to the predicted token distribution. The generated embedding is then fed back as input for another forward pass. We show that the model can learn to ponder in this way through self-supervised learning, without any human annotations. Experiments across three widely used open-source architectures—GPT-2, Pythia, and LLaMA—and extensive downstream task evaluations demonstrate the effectiveness and generality of our method. On 9 downstream benchmarks, our pondering-enhanced Pythia models significantly outperform the official Pythia models. Notably, our PonderPythia models demonstrate remarkable effectiveness: PonderPythia-2.8B surpasses Pythia-6.9B and rivals Pythia-12B, while our PonderPythia-1B matches TinyLlama-1.1B, a model trained on 10 times more data.

Vision-Language Models (VLMs), such as CLIP, have shown strong zero-shot generalization but remain highly vulnerable to adversarial perturbations, posing serious risks in real-world applications. Test-time defenses for VLMs have recently emerged as a promising and efficient approach to defend against adversarial attacks without requiring costly large-scale retraining. In this work, we uncover a surprising phenomenon: under diverse input transformations, adversarial images in CLIP’s feature space consistently shift along a dominant direction, in contrast to the dispersed patterns of clean images. We hypothesize that this dominant shift, termed the Defense Direction, opposes the adversarial shift, pointing features back toward their correct class centers. Building on this insight, we propose Directional Bias-guided Defense (DBD), a test-time framework that estimates the Defense Direction and employs a DB-score–based two-stream reconstruction strategy to recover robust representations. Experiments on 15 datasets demonstrate that DBD not only achieves SOTA adversarial robustness while preserving clean accuracy, but also reveals the counterintuitive result that adversarial accuracy can even surpass clean accuracy. This demonstrates that adversarial perturbations inherently encode directional priors about the true decision boundary.

While search-augmented large language models (LLMs) exhibit impressive capabilities, their reliability in complex multi-hop reasoning remains limited. This limitation arises from three fundamental challenges: decomposition errors, where tasks are incorrectly broken down; retrieval missing, where key evidence fails to be retrieved; and reasoning errors, where flawed logic propagates through the reasoning chain. A single failure in any of these stages can derail the final answer. We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process. ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain. This targeted correction mechanism turns brittle reasoning into a more resilient process. Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle, with the 3B model achieving +8.48% EM and +11.56% F1, and the 7B model achieving +5.38% EM and +7.22% F1 over previous state-of-the-art(SOTA) results. These findings suggest that erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs.

Recent generative pre-trained vision–language (GPTv) models have achieved remarkable success in multi-modal understanding, inspiring their adaptation to medical imaging tasks such as disease diagnosis and visual question answering (VQA). However, current instruction-tuned GPTv models suffer from two key challenges: (1) medical attributes (e.g., disease names, severity grades) are encoded as plain text tokens, collapsing semantically distinct concepts into nearly identical textual sequences; and (2) inadequate textual supervision weakens visual representation learning, leading to severe inter-attribute confusion and misaligned vision–language embeddings. To address these limitations, we introduce attribute tokens (AttTok), a set of pre‑defined special tokens that uniquely encode clinical attributes (e.g., imaging modality, diagnosis, severity) within a structured token space. Complemented by attribute‑centric embedding books, AttTok serves as anchor points for aligning both visual and textual modalities into a shared, discriminative representation space. Building on this foundation, we design two key components: an attribute‑centric cross attention (ACC) adapter, which breaks the vision‑to‑text information‑flow bottleneck and enriches the visual encoder with discriminative attribute knowledge, and an attribute‑centric matching (ACM) loss, which enforces robust multi‑modal alignment centered on the attribute tokens. Extensive experiments on five medical classification benchmarks and three VQA datasets demonstrate that AttTok substantially improves both discriminative accuracy and medical knowledge reasoning, establishing a new paradigm for medical GPTv models with clinically discriminative understanding.

Monocular 3D reconstruction methods and vision-language models (VLMs) demonstrate impressive results on standard benchmarks, yet their true understanding of geometric properties remains unclear. We introduce GIQ, a comprehensive benchmark specifically designed to evaluate the geometric reasoning capabilities of vision and vision-language foundation models. GIQ comprises synthetic and real-world images and corresponding 3D meshes of diverse polyhedra—including Platonic, Archimedean, Johnson, and Catalan solids, as well as stellations and compound shapes—covering varying levels of complexity and symmetry. Through systematic experiments involving monocular 3D reconstruction, 3D symmetry detection, mental rotation tests, and zero-shot shape classification tasks, we reveal significant shortcomings in current models. State-of-the-art reconstruction algorithms trained on extensive 3D datasets struggle to reconstruct even basic geometric forms accurately. While foundation models effectively detect specific 3D symmetry elements via non-linear probing, they falter significantly in tasks requiring detailed geometric differentiation, such as mental rotation. Moreover, advanced vision-language assistants exhibit remarkably low accuracy on complex polyhedra, systematically misinterpreting basic properties like face geometry, convexity, and compound structures. GIQ is publicly available at toomanymatts.github.io/giq-benchmark/, providing a structured platform to benchmark critical gaps in geometric intelligence and facilitate future progress in robust, geometry-aware representation learning.

The rapid rise of image generation calls for detection methods that are both interpretable and reliable. Existing approaches, though accurate, act as black boxes and fail to generalize to out-of-distribution data, while multi-modal large language models (MLLMs) provide reasoning ability but often hallucinate. To address these issues, we construct \textbf{FakeXplained} dataset of AI-generated images annotated with bounding boxes and descriptive captions that highlight synthesis artifacts, forming the basis for human-aligned, visually grounded reasoning. Leveraging \textbf{FakeXplained}, we develop \textbf{FakeXplainer} which fine-tunes MLLMs with a progressive training pipeline, enabling accurate detection, artifact localization, and coherent textual explanations. Extensive experiments show that \textbf{FakeXplainer} not only sets a new state-of-the-art in detection and localization accuracy ($98.2\%$ accuracy, $36.0\%$ IoU), but also demonstrates strong robustness and out-of-distribution generalization, uniquely delivering spatially grounded, human-aligned rationales. The code and dataset are available at: \href{https://github.com/Gennadiyev/FakeXplain}{https://github.com/Gennadiyev/FakeXplain}.

DeepSeek-R1-Zero has successfully demonstrated the emergence of reasoning capabilities in LLMs purely through Reinforcement Learning (RL). Inspired by this breakthrough, we explore how RL can be utilized to enhance the reasoning capability of MLLMs. However, direct training with RL struggles to activate complex reasoning capabilities such as questioning and reflection in MLLMs, due to the absence of substantial high-quality multimodal reasoning data. To address this issue, we propose the reasoning MLLM, Vision-R1, to improve multimodal reasoning capability. Specifically, we first construct a high-quality multimodal CoT dataset without human annotations by leveraging an existing MLLM and DeepSeek-R1 through modality bridging and data filtering to obtain a 200K multimodal CoT dataset, Vision-R1-cold dataset. It serves as cold-start initialization data for Vision-R1. To mitigate the optimization challenges caused by overthinking after cold start, we propose Progressive Thinking Suppression Training (PTST) strategy and employ Group Relative Policy Optimization (GRPO) with the hard formatting result reward function to gradually refine the model's ability to learn correct and complex reasoning processes on the multimodal math dataset. Comprehensive experiments show our model achieves an average improvement of $\sim$6\% across various multimodal math reasoning benchmarks using only a 10K multimodal math data during RL training. Vision-R1-7B achieves a 73.5\% accuracy on the widely used MathVista benchmark, which is only 0.4\% lower than the leading reasoning model, OpenAI O1. Scaling up the amount of multimodal math data in the RL training, Vision-R1-32B and Vison-R1-72B achieves 76.4\% and 78.2\% MathVista benchmark scores, respectively.