ICLR 2026 Papers

Scaling Vision-Language-Action (VLA) models on large-scale data offers a promising path to achieving a more generalized driving intelligence. However, VLA models are limited by a ``supervision deficit'': the vast model capacity is supervised by sparse, low-dimensional actions, leaving much of their representational power underutilized. To remedy this, we propose DriveVLA-W0, a training paradigm that employs world modeling to predict future images. This task generates a dense, self-supervised signal that compels the model to learn the underlying dynamics of the driving environment. We showcase the paradigm's versatility by instantiating it for two dominant VLA archetypes: an autoregressive world model for VLAs that use discrete visual tokens, and a diffusion world model for those operating on continuous visual features. Building on the rich representations learned from world modeling, we introduce a lightweight action expert to address the inference latency for real-time deployment. Extensive experiments on the NAVSIM benchmark and a 700x larger in-house dataset demonstrate that DriveVLA-W0 significantly outperforms BEV and VLA baselines. Crucially, it amplifies the data scaling law, showing that performance gains accelerate as the training dataset size increases. Code will be made publicly available.

Large Language Models (LLMs), despite being trained on text alone, surprisingly develop rich visual priors. These priors allow latent visual capabilities to be unlocked for vision tasks with a relatively small amount of multimodal data, and to perform symbolic visual generation tasks without ever having seen an image. Through systematic analysis, we reveal that visual priors—the implicit, emergent knowledge about the visual world acquired during language pre-training—are composed of separable perception and reasoning priors with unique scaling trends and origins. We show that an LLM's latent visual reasoning ability is predominantly developed by pre-training on reasoning-centric data (\eg, code, math, academia) and scales progressively. This reasoning prior acquired from language pre-training is transferable and universally applicable to visual reasoning. In contrast, the perception prior emerges more diffusely from broad corpora, and perception ability is more sensitive to the vision encoder and visual instruction tuning data. In parallel, text describing the visual world proves crucial, though its performance impact saturates rapidly. Leveraging these insights, we propose a data-centric recipe for pre-training vision-aware LLMs and verify it in 1T token scale pre-training. Our findings are grounded in over 100 controlled experiments consuming 500,000 GPU-hours, spanning the full MLLM construction pipeline—from LLM pre-training to visual alignment and supervised multimodal fine-tuning—across five model scales, a wide range of data categories and mixtures, and multiple adaptation setups. Along with our main findings, we also propose and investigate several hypotheses, and introduce a Multi-Level Existence Bench (MLE-Bench) to facilitate future research. Together, this work provides a new way of deliberately cultivating visual priors from language pre-training, paving the way for the next generation of multimodal LLMs. We recommend a visit to our anonymous project page (https://anonymouspaperweb.github.io/lsbs/) for an interactive reading.

Geometric problem solving constitutes a critical branch of mathematical reasoning, requiring precise analysis of shapes and spatial relationships. Current evaluations of geometric reasoning in vision-language models (VLMs) face limitations, including the risk of test data contamination from textbook-based benchmarks, overemphasis on final answers over reasoning processes, and insufficient diagnostic granularity. To address these issues, we present GeoBench, a hierarchical benchmark featuring four reasoning levels in geometric problem-solving: Visual Perception, Goal-Oriented Planning, Rigorous Theorem Application, and Self-Reflective Backtracking. Through six formally verified tasks generated via TrustGeoGen, we systematically assess capabilities ranging from attribute extraction to logical error correction. Experiments reveal that while reasoning models like OpenAI-o3 outperform general MLLMs, performance declines significantly with increasing task complexity. Key findings demonstrate that sub-goal decomposition and irrelevant premise filtering critically influence final problem-solving accuracy, whereas Chain-of-Thought prompting unexpectedly degrades performance in some tasks. These findings establish GeoBench as a comprehensive benchmark while offering actionable guidelines for developing geometric problem-solving systems. Our benchmark and code are released at https://anonymous.4open.science/r/GeoBench-4BF4.

Calibrated confidence estimates are necessary for large language model (LLM) outputs to be trusted by human users. While LLMs can express their confidence in human-interpretable ways, verbalized LLM-generated confidence scores have empirically been found to be miscalibrated, reporting high confidence on instances with low accuracy and thereby harming trust and safety. We hypothesize that this overconfidence often stems from a given LLM’s heightened suggestibility when faced with claims that it encodes little information about; we empirically validate this hypothesis, finding more suggestibility on lower-accuracy claims. Building on this finding, we introduce Distractor-Normalized Coherence (DINCO), which estimates and accounts for an LLM’s suggestibility bias by having the model verbalize its confidence independently across several self-generated distractors (i.e. alternative claims), and normalizes by the total verbalized confidence. To further improve calibration, we leverage generator-validator disagreement, augmenting normalized validator confidence with a consistency-based estimate of generator confidence. Here, we frame the popular approach of self-consistency as leveraging coherence across sampled generations, and normalized verbalized confidence as leveraging coherence across validations on incompatible claims, allowing us to integrate these complementary dimensions of coherence into DINCO. Moreover, our analysis shows that DINCO provides less saturated, and therefore more usable, confidence estimates, and that further sampling alone cannot close the gap between DINCO and baselines, with DINCO at 10 runs outperforming self-consistency at 100. We include our code in the supplementary.

Deploying clinical ML is slow and brittle: models that work at one hospital often degrade under distribution shifts at the next. In this work, we study a simple question -- can large language models (LLMs) create portable patient embeddings i.e. representations of patients enable a downstream predictor built on one hospital to be elsewhere with minimal-to-no retraining and fine-tuning. To do so, we map from irregular ICU time series onto concise natural language summaries using a frozen LLM, then embed each summary with a frozen text embedding model to obtain a fixed length vector capable of serving as input to a variety of downstream predictors. Across three cohorts (MIMIC-IV, HIRID, PPICU), on multiple clinically grounded forecasting and classification tasks, we find that our approach is simple, easy to use and surprisingly competitive with in-distribution with grid imputation, self-supervised representation learning, and time series foundation models, while exhibiting smaller relative performance drops when transferring to new hospitals. We study the variation in performance across prompt design, with structured prompts being crucial to reducing the variance of the predictive models without altering mean accuracy. We find that using these portable representations improves few-shot learning and does not increase demographic recoverability of age or sex relative to baselines, suggesting little additional privacy risk. Our work points to the potential that LLMs hold as tools to enable the scalable deployment of production grade predictive models by reducing the engineering overhead.

Large language models (LLMs) often solve drill-style math exercises yet fail to apply the concept right when the problem requires genuine understanding. Popular outcome-based RL pipelines reinforce final answers but provide little fine-grained conceptual signal, so models improve at pattern reuse rather than concept selection and use. We introduce $\textit{CORE}$ (Concept-Oriented REinforcement), an algorithm-agnostic training framework that turns explicit concepts into a controllable supervision signal. Starting from a high-quality, low-contamination textbook resource that links verifiable exercises to concise concept descriptions, we run a sanity probe showing LLMs can restate definitions while missing concept-linked quizzes, quantifying the conceptual reasoning gap. $\textit{CORE}$ then (i) synthesizes additional concept-aligned quizzes, (ii) injects concept snippets into rollouts, and (iii) reinforces trajectories that correctly apply the injected concept while constraining drift with a lightweight divergence penalty; the procedure is compatible with standard policy-gradient methods (e.g., GRPO). On a 7B-class model, $\textit{CORE}$ yields consistent gains over a vanilla baseline and reinforcement-only training across in-domain concept–exercise suites and diverse out-of-domain math benchmarks (GSM8K, SVAMP, MAWPS, SAT-Math, OlympiadBench, Gaokao, Minerva-Math, CounterMath, TheoremQA). Improvements are largest on concept-heavy categories while maintaining or modestly improving drill performance. $\textit{CORE}$ demonstrates that concept-injected, outcome-regularized rollouts supply the missing fine-grained supervision needed to bridge drill competence and true conceptual reasoning—without committing to a particular RL algorithm or certain process-based verifiers.

Reinforcement Learning (RL) has become the _de facto_ standard for tuning LLMs to solve tasks involving reasoning. However, growing evidence shows that such models often suffer from a significant loss in diversity. We argue that this arises because RL implicitly optimizes the Reverse KL to a target distribution, which concentrates on certain high-probability regions of the target while neglecting others. In this work, we instead begin from an explicit target distribution, obtained by filtering out incorrect answers while preserving the relative probabilities of correct ones. Starting from a pre-trained LLM, we approximate this target distribution using Amari's $\alpha$-divergence family, which unifies prior approaches and enables direct control of the precision–diversity trade-off by interpolating between mode-seeking and mass-covering divergences. On a Lean theorem-proving benchmark, our method achieves state-of-the-art performance along the coverage–precision Pareto frontier, unmatched by other methods along the coverage axis.

Understanding human behavior requires measuring behavioral actions. Due to its complexity, behavior is best mapped onto a rich, semantic structure such as language. Emerging multimodal large language models (MLLMs) are promising candidates, but their fine-grained action understanding ability has not been fully examined. In this work, we reformulate EPIC-KITCHENS-100, one of the largest and most challenging egocentric action recognition datasets, into a MLLM benchmark (EPIC-KITCHENS-100-MQA). We show that when we sample difficult answers based on specialist models as distractors, leading MLLMs struggle to recognize the correct actions. How can we increase the performance of MLLMs? We curated a supervised finetuning dataset that includes `hard' action recognition, temporal detection, captioning, and free-form question answering to improve models' diverse action understanding capabilities. We introduce a new model called LLaVAction that adds an action token to boost models' attention on visual tokens and a two-stage pipeline to obtain structured actions. LLaVAction greatly improves the MLLMs' ability of action understanding, achieving strong improvements on both MLLM benchmarks (21 points in accuracy over GPT-4o on EPIC-KITCHENS-100-MQA) and action recognition benchmarks, suggesting that our methods prepare MLLMs to be a promising path forward for complex action tasks. Code, data, benchmark and models will be available upon acceptance.

Large Language Models (LLMs) are increasingly applied to complex tasks that require extended reasoning. In such settings, models often benefit from diverse chains-of-thought to arrive at multiple candidate solutions. This requires two competing objectives: to inject enough stochasticity to explore multiple reasoning chains, and to ensure sufficient accuracy and quality in each path. Existing works pursue the first objective by increasing exploration at highly uncertain steps with higher temperature or larger candidate token sets, while others improve reliability by rejecting samples with low confidence post generation, implying that low confidence correlates with low answer quality. These two lines of thought are in conflict, as they conflate different sources of uncertainty. To resolve this, we argue that the decoding rule should be calibrated by *correctness*, not confidence alone. We should sample from tokens with higher estimated correctness, and reduce sampling where expected correctness is low. We propose simple strategies that achieve this goal: **Greedy-Threshold** makes sampling greedy at very low confidence steps. **Calibrated-TopK** and **Calibrated-ε** set truncation threshold based on estimated rank-wise correctness. Together, our findings challenge prevailing heuristics about decoding under uncertainty, showing consistent gains across math and general reasoning benchmarks.

Large Language Models (LLMs) have become foundational tools in natural language processing, powering a wide range of applications and research. Many studies have shown that LLMs share significant similarities with the human brain. Neuroscience research has found that a small subset of biological neurons in the human brain are crucial for core cognitive functions, which raises a fundamental question: do LLMs also contain a small subset of critical neurons? In this paper, we investigate this question by proposing a Perturbation-based Causal Identification of Critical Neurons method to systematically locate such critical neurons in LLMs. Our findings reveal three key insights: (1) LLMs contain ultra-sparse critical neuron sets. Disrupting these critical neurons can cause a 72B-parameter model with over 1.1 billion neurons to completely collapse, with perplexity increasing by up to 20 orders of magnitude; (2) These critical neurons are not uniformly distributed, but tend to concentrate in the outer layers, particularly within the MLP down\_proj components; (3) Performance degradation exhibits sharp phase transitions, rather than a gradual decline, when these critical neurons are disrupted. Through comprehensive experiments across diverse model architectures and scales, we provide deeper analysis of these phenomena and their implications for LLM robustness and interpretability. These findings can offer guidance for developing more robust model architectures and improving deployment security in safety-critical applications. Our code is available at https://github.com/qqqqqqqzx/The-Achilles-Heel-of-LLMs.

Vision-and-Language Navigation (VLN) agents have shown strong capabilities in following natural language instructions. However, they often struggle to generalize across environments due to catastrophic forgetting, which limits their practical use in real-world settings where agents must continually adapt to new domains. We argue that overcoming forgetting across environments hinges on decoupling global scene reasoning from local perceptual alignment, allowing the agent to adapt to new domains while preserving specialized capabilities. To this end, we propose M$^3$E, the Mixture of Macro and Micro Experts, an environment-aware hierarchical MoE framework for continual VLN. Our method introduces a dual-router architecture that separates navigation into two levels of reasoning. A macro-level, scene-aware router selects strategy experts based on global environmental features (e.g., office vs. residential), while a micro-level, instance-aware router activates perception experts based on local instruction-vision alignment for step-wise decision making. To preserve knowledge across domains, we adopt a dynamic momentum update strategy that identifies expert utility in new environments and selectively updates or freezes their parameters. We evaluate M$^3$E in a domain-incremental setting on the R2R and REVERIE datasets, where agents learn across unseen scenes without revisiting prior data. Results show that our method consistently outperforms standard fine-tuning and existing continual learning baselines in both adaptability and knowledge retention, offering a parameter-efficient solution for building generalizable embodied agents.

This paper presents Omni-View, which extends the unified multimodal understanding and generation to 3D scenes based on multiview images, exploring the principle that ``generation facilitates understanding". Consisting of understanding model, texture module, and geometry module, Omni-View jointly models scene understanding, novel view synthesis, and geometry estimation, enabling synergistic interaction between 3D scene understanding and generation tasks. By design, it leverages the spatiotemporal modeling capabilities of its texture module responsible for appearance synthesis, alongside the explicit geometric constraints provided by its dedicated geometry module, thereby enriching the model’s holistic understanding of 3D scenes. Trained with a two-stage strategy, Omni-View achieves a state-of-the-art score of 55.4 on the VSI-Bench benchmark, outperforming existing specialized 3D understanding models, while simultaneously delivering strong performance in both novel view synthesis and 3D scene generation.

The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications. In this work, we introduce **VLM-SubtleBench**, a benchmark designed to evaluate VLMs on *subtle comparative reasoning*. Our benchmark covers ten difference types—Attribute, State, Emotion, Temporal, Spatial, Existence, Quantity, Quality, Viewpoint, and Action—and curate paired question–image sets reflecting these fine-grained variations. Unlike prior benchmarks restricted to natural image datasets, our benchmark spans diverse domains, including industrial, aerial, and medical imagery. Through extensive evaluation of both proprietary and open-source VLMs, we reveal systematic gaps between model and human performance across difference types and domains, and provide controlled analyses highlighting where VLMs’ reasoning sharply deteriorates. Together, our benchmark and findings establish a foundation for advancing VLMs toward human-level comparative reasoning.

Reinforcement learning (RL) has shown strong potential for enhancing reasoning in multimodal large language models (MLLMs), yet existing video reasoning methods often rely on coarse sequence-level rewards or single-factor token selection. Such approaches neglect fine-grained links among visual inputs, temporal dynamics, and linguistic outputs, limiting both accuracy and interpretability. We propose Video-KTR, a modality-aware policy shaping framework that performs selective, token-level RL by combining three attribution signals: (1) visual-aware tokens identified via counterfactual masking to reveal perceptual dependence; (2) temporal-aware tokens detected through frame shuffling to expose causal and temporal sensitivity; and (3) high-entropy tokens signaling predictive uncertainty. By reinforcing only the union of key tokens, Video-KTR focuses learning on semantically informative, modality-sensitive content while filtering out low-value tokens. Across five challenging benchmarks, Video-KTR achieves state-of-the-art or highly competitive results—42.7% on Video-Holmes, surpassing GPT-4o—with consistent gains on both reasoning-centric and general video understanding tasks. Ablation studies verify the complementary roles of the attribution signals and the robustness of targeted token-level updates. Overall, Video-KTR improves accuracy and interpretability, offering a simple, drop-in extension to RL for complex video reasoning.

Mean flow (MeanFlow) enables efficient, high-fidelity image generation, yet its single-function evaluation (1-NFE) generation often cannot yield compelling results. We address this issue by introducing RMFlow, an efficient multimodal generative model that integrates a coarse 1-NFE MeanFlow transport with a subsequent tailored noise-injection refinement step. RMFlow approximates the average velocity of the flow path using a neural network trained with a new loss function that balances minimizing the Wasserstein distance between probability paths and maximizing sample likelihood. RMFlow achieves near state-of-the-art results on text-to-image, context-to-molecule, and time-series generation using only 1-NFE, at a computational cost comparable to the baseline MeanFlows.

Pretrained medical foundation models (FMs) have shown strong generalization across diverse imaging tasks, such as disease classification in radiology and tumor grading in histopathology. While recent advances in parameter-efficient finetuning have enabled effective adaptation of FMs to downstream tasks, these approaches are typically designed for a single modality. In contrast, many clinical workflows rely on joint diagnosis from heterogeneous domains, such as radiology and pathology, where fully leveraging the representation capacity of multiple FMs remains an open challenge. To address this gap, we propose Concept Tuning and Fusing (CTF), a parameter-efficient framework that uses clinically grounded concepts as a shared semantic interface to enable cross-modal co-adaptation before fusion. By incorporating task-specific concepts that are relevant across modalities, CTF aligns radiology and pathology representations, thereby enhancing their complementarity and enabling interpretation. We further design a Global–Context–Shared Prompt (GCSP) mechanism, which employs a small set of learnable tokens to capture domain-specific priors, shared patient-level information, and cross-domain context. The resulting concept alignment scores from each modality are then fused to produce a final prediction. Extensive experiments demonstrate that CTF outperforms strong unimodal, latent-fusion, and adapter-based baselines (e.g., AUC 0.903 on TCGA-GBMLGG). Notably, CTF achieves these gains without finetuning the full FMs, requiring only 0.15\% additional parameters, thus highlighting the effectiveness of concept-based multimodal co-adaptation. Our code is anonymously available at: https://anonymous.4open.science/r/CTF-27C2.

Multimodal learning is crucial for ecological applications, which rely on heterogeneous data sources (e.g., satellite imagery, environmental time series, tabular predictors, bioacoustics) but often suffer from incomplete data across and within modalities (e.g., unavailable satellite image due to cloud cover, missing records in a time series). While data masking strategies have been used to improve robustness to missing data by exposing models to varying input subsets during training, existing approaches typically rely on static masking and inadequately explore the space of input combinations. As a result, they fail to address modality imbalance, a critical challenge in multimodal learning where dominant modalities hinder the optimization of others. To fill this gap, we introduce Modality Imbalance-Aware Masking (MIAM), a dynamic masking strategy that: (i) explores the full space of input combinations; (ii) prioritizes informative or challenging subsets; and (iii) adaptively increases the masking probability of dominant modalities based on their relative performance and learning dynamics. We evaluate MIAM on two key ecological datasets, GeoPlant and TaxaBench, with diverse modality configurations, and show that MIAM significantly improves robustness and predictive performance over previous masking strategies. In addition, MIAM supports fine-grained contribution analysis across and within modalities, revealing which variables, time segments, or image regions most strongly drive performance.

Video anomaly detection (VAD) aims to identify abnormal events in videos. Traditional VAD methods generally suffer from the high costs of labeled data and full training, thus some recent works have explored leveraging frozen multi-modal large language models (MLLMs) in a tuning-free manner to perform VAD. However, their performance is limited as they directly inherit pre-training biases and cannot adapt internal representations to specific video contexts, leading to difficulties in handling subtle or ambiguous anomalies. To address these limitations, we propose a novel intervention framework, termed SteerVAD, which advances MLLM-based VAD by shifting from passively reading to actively steering and rectifying internal representations. Our approach first leverages the gradient-free representational separability analysis (RSA) to identify top attention heads as latent anomaly experts (LAEs) which are most discriminative for VAD. Then a hierarchical meta-controller (HMC) generates dynamic rectification signals by jointly conditioning on global context and these LAE outputs. The signals execute targeted, anisotropic scaling directly upon the LAE representation manifolds, amplifying anomaly-relevant dimensions while suppressing inherent biases. Extensive experiments on mainstream benchmarks demonstrate our method achieves state-of-the-art performance among tuning-free approaches requiring only 1\% of training data, establishing it as a powerful new direction for video anomaly detection. The code will be released upon the publication.

Existing large language model (LLM)-based embeddings typically adopt an encoder-only paradigm, treating LLMs as static feature extractors and overlooking their core gener- ative strengths. We introduce GIRCSE (Generative Iterative Refinement for Contrastive Sentence Embeddings), a novel framework that leverages autoregressive generation to iter- atively refine semantic representations. By producing sequences of soft tokens optimized under a contrastive objective, GIRCSE captures latent concepts and implicit semantics that encoder-only methods often miss. To guide this process, we propose an Iterative Contrastive Refinement (ICR) objective that encourages each refinement step to yield bet- ter representations. Extensive experiments show that GIRCSE outperforms strong LLM- based embedding baselines on the MTEB embedding benchmark. Moreover, GIRCSE ex- hibits an emergent test-time scaling property: generating more tokens at inference steadily improves embedding quality. Our results establish generative iterative refinement as a new paradigm for representation learning.