ICLR 2026 Papers

While Large Vision-Language Models (LVLMs) have achieved substantial progress in video understanding, their application to long video reasoning is hindered by uniform frame sampling and static textual reasoning, which are inefficient and struggle to handle visually intensive video tasks. To overcome these challenges, in this paper, we introduce the concept of thinking with long videos and propose a novel framework FrameThinker. Within this framework, LVLMs are able to iteratively interrogate video content. Developing such video reasoning capabilities in LVLMs presents notable challenges, particularly in adapting the model to new video actions (e.g. select frame), and designing reward functions to guide LVLMs to adopt the newly introduced action. To solve these challenges, we propose a two-phase training strategy, first employing Supervised Fine-Tuning (SFT) to instill fundamental action capabilities, followed by Reinforcement Learning (RL) to optimize a strategic decision-making policy. Notably, in this RL phase, we conduct an in-depth and comprehensive exploration of the reward design for each action and format reward. Extensive experiments on reasoning benchmarks like Video-Holmes, LongVideo-Reason, and long-video understanding benchmarks such as LongVideoBench, MLVU, VideoMME, and LVBench, demonstrate that FrameThinker gets a significant average improvement of +10.4\% over baselines while drastically reducing the number of processed frames. Most notably, our 7B model, FrameThinker establishes a new state-of-the-art on LongVideo-Reason, achieving 76.1\% accuracy using an average of only 20.6 frames. This not only outperforms the competitive LongVILA-R1 (72.0\%) but does so with over 20x fewer frames (vs. 512), demonstrating unparalleled efficiency and effectiveness. Our code is available at: \url{https://anonymous.4open.science/r/FrameThinker-B5FD}.

Computed Tomography (CT) is one of the most widely used and diagnostically information-dense imaging modalities, covering critical organs such as the heart, lungs, liver, and colon. Clinical interpretation relies on both \textbf{slice-driven} local features (e.g., sub-centimeter nodules, lesion boundaries) and \textbf{volume-driven} spatial representations (e.g., tumor infiltration, inter-organ anatomical relations). However, existing Large Vision–Language Models (LVLMs) remain fragmented in CT slice versus volumetric understanding: slice-driven LVLMs show strong generalization but lack cross-slice spatial consistency, while volume-driven LVLMs explicitly capture volumetric semantics but suffer from coarse granularity and poor compatibility with slice inputs. The absence of a unified modeling paradigm constitutes a major bottleneck for the clinical translation of medical LVLMs. We present \textbf{OmniCT}, a powerful unified slice–volume LVLM for CT scans, which makes three contributions: \textbf{(i) Spatial Consistency Enhancement (SCE):} volumetric slice composition combined with tri-axial positional encoding introduces volumetric consistency, and an MoE hybird projection enables efficient slice–volume adaptation; \textbf{(ii) Organ-level Semantic Enhancement (OSE):} segmentation and ROI localization explicitly align anatomical regions, emphasizing lesion- and organ-level semantics; \textbf{(iii) MedEval-CT:} the largest slice–volume CT dataset and hybrid benchmark integrates multi-level metrics for unified evaluation. OmniCT consistently outperforms existing methods with a substantial margin across diverse clinical tasks, satisfies both micro-level detail sensitivity and macro-level spatial reasoning, and establishes a new paradigm for cross-dimensional medical imaging modeling. Our project is available at \href{https://anonymous.4open.science/r/OmniCT}{link}.

Top-down transparency typically analyzes language model activations using probes with scalar or single-token outputs, limiting the range of behaviors that can be captured. To alleviate this issue, we develop a more expressive probe that can directly output natural language and perform LatentQA: the task of answering open-ended questions about activations. A key difficulty in developing such a probe is collecting a dataset mapping activations to natural-language descriptions. In response, we propose an approach for generating a pseudo-labeled dataset of activations and associated question-answer pairs and develop a fine-tuning method for training a decoder LLM on this dataset. We then validate our decoder’s fidelity by assessing its ability to read and steer model activations. First, we evaluate the decoder on a number of supervised reading tasks with a known answer, such as uncovering hidden system prompts and relational knowledge extraction, and observe that it outperforms competitive probing baselines. Second, we demonstrate that the decoder is precise enough to steer the target model to exhibit behaviors unseen during training. Finally, we show that LatentQA scales well with increasing dataset and model size, which is promising given how easily our approach can generate additional pseudo-labels.

Vision-Language Models (VLMs) for radiology report generation are typically trained to mimic the narrative flow of human experts. However, we identify a potential limitation in this conventional paradigm. We hypothesize that optimizing for narrative coherence encourages models to rely on linguistic priors and inter-sentence correlations, which can weaken their grounding in direct visual evidence and lead to factual inaccuracies. To investigate this, we design a controlled experiment demonstrating that as textual context increases, a model's reliance on the input image systematically decays. We propose LLaVA-TA (Topic-guided and Anatomy-aware), a new fine-tuning framework that directly addresses this challenge by re-engineering the generation process. Instead of producing a linear narrative, LLaVA-TA decomposes the report into a set of independent, clinically-relevant topics. By training the model to generate a discrete finding for each topic conditioned on both the full image and its corresponding anatomical region, we reduce the model's reliance on narrative flow and enforce stricter visual grounding. Our experiments show that LLaVA-TA sets a new state of the art on the MIMIC-CXR dataset, significantly improving clinical accuracy on metrics like RadGraph F1 (from 29.4 to 44.0) and CheXpert F1-14 (from 39.5 to 71.5) over strong baselines. Our work demonstrates that dismantling a report's narrative structure to enforce independent, visually-grounded observations is a crucial and effective step toward building more accurate and reliable medical VLMs.

While augmenting Multimodal Large Language Models (MLLMs) with tools is a promising direction, current approaches face critical limitations. They often rely on single, atomic tools, failing to address the challenges of multi-turn planning, and they do not equip models with the ability to select effective tool combinations for complex tasks. To overcome these limitations, we introduce AdaReasoner, a framework that teaches models to perform dynamic tool orchestration for iterative visual reasoning. Our paradigm is designed to support a broad spectrum of tools, including computationally intensive, expert-model-based services. It features a comprehensive design that includes a new data curation methodology and a tailored Tool GRPO algorithm to optimize multi-turn tool-calling trajectories, which yields state-of-the-art models that achieve substantial gains over their baselines (+38.7\% average on 7B) and reach near-perfect accuracy on complex benchmarks like Visual Spatial Planning (97.6\%). This performance surpasses leading proprietary systems such as GPT-5 and Claude Sonnet 4, demonstrating that our approach can effectively overcome scale-based limitations by augmenting smaller models with powerful tool-use capabilities. Critically, we find that AdaReasoner develops emergent, self-adaptive behaviors: it learns to autonomously adopt beneficial tools, discard irrelevant ones, and modulate its usage frequency. This ability to curate its own optimal problem-solving strategies represents a significant step toward building more robust, scalable, and reliable reasoning agents.

Large Language Models (LLMs) have shown strong performance on text-attributed graphs (TAGs) due to their superior semantic understanding ability on textual node features. However, their effectiveness as predictors in the few-shot semi-supervised setting, where labeled nodes are rather limited, remains constrained since fine-tuning LLMs usually requires sufficient labeled data, especially when the TAG shows complex structural patterns. In essence, this paper targets two key challenges: (i) the difficulty of generating and selecting reliable pseudo labels on TAGs for LLMs, and (ii) the need to mitigate potential label noise when fine-tuning LLMs with pseudo labels. To counter the challenges, we propose a new framework, GNN-as-Judge, which can unleash the power of LLMs for few-shot semi-supervised learning on TAGs by incorporating the structural inductive bias of Graph Neural Networks (GNNs). Specifically, GNN-as-Judge introduces a collaborative pseudo-labeling strategy that first identifies the most influenced unlabeled nodes from labeled nodes, then exploits both the agreement and disagreement patterns between LLMs and GNNs to generate reliable labels. Furthermore, we develop a weakly-supervised LLM fine-tuning algorithm that can distill the knowledge from informative pseudo labels while mitigating the potential label noise. Experiments on different TAG datasets demonstrate that GNN-as-Judge significantly outperforms existing methods, especially under low-resource regimes.

Recent advances in image generation, often driven by proprietary systems like GPT-4o Image Gen, regularly introduce new capabilities that reshape how users interact with these models. Existing benchmarks often lag behind and fail to capture these emerging use cases, leaving a gap between community perceptions of progress and formal evaluation. To address this, we present ECHO, a framework for constructing benchmarks directly from real-world evidence of model use: social media posts that showcase novel prompts and qualitative user judgments. Applying this framework to GPT-4o Image Gen, we construct a dataset of over 35,000 prompts curated from such posts. Our analysis shows that ECHO (1) discovers creative and complex tasks absent from existing benchmarks, such as re-rendering product labels across languages or generating receipts with specified totals, (2) more clearly distinguishes state-of-the-art models from alternatives, and (3) surfaces community feedback that we use to inform the design of metrics for model quality (e.g., measuring observed shifts in color, identity, and structure). Our website is at https://echo-bench.github.io.

Large language models have revolutionized natural language processing, yet their deployment remains hampered by the substantial memory and runtime overhead of the transformer’s Key-Value cache. To mitigate this, recent methods employ a scoring-aggregation framework to evict unimportant cache entries, based on the "stability assumption"—that a fixed subset of entries remains consistently important during generation. However, prior work has largely focused on refining importance indicators for scoring, while defaulting to mean aggregation due to a faithful trust in the stability assumption. In this work, we argue that this underlying assumption is inherently fragile, making mean aggregation highly vulnerable in extreme cases. To counter this, we propose a simple yet elegant defensive aggregation strategy: a two-step, linear-time approach that controls worst-case risk, thereby defending against extreme cases with negligible computational overhead. Embodying this strategy, we propose a novel cache eviction method, DefensiveKV and its extension, Layer-DefensiveKV, which incorporates layer-wise budget allocation. Across seven task domains (18 datasets), our methods reduce generation quality loss by 2.3× and 4.3× respectively, versus the strongest baseline under a 20\% cache size. These results set new performance benchmarks and pioneer a promising direction for optimizing cache eviction against underlying fragility through worst-case risk management.

The recent development of foundation models for time series data has generated considerable interest in using such models across a variety of applications. Although foundation models achieve state-of-the-art predictive performance, their calibration properties remain relatively underexplored, despite the fact that calibration can be critical for many practical applications. In this paper, we investigate the calibration-related properties of five recent time series foundation models and two competitive baselines. We perform a series of systematic evaluations assessing model calibration (i.e., over- or under-confidence), effects of varying prediction heads, and calibration under long-term autoregressive forecasting. We find that time series foundation models are consistently better calibrated than baseline models and tend not to be either systematically over- or under-confident, in contrast to the overconfidence often seen in other deep learning models.

Tool design and use reflect the ability to understand and manipulate the physical world through creativity, planning, and foresight. As such, it is often regarded as a measurable indicator of cognitive intelligence across biological species. While much of today’s research on robotics intelligence focuses on generating better control strategies, inventing smarter tools offers a complementary form of physical intelligence: moving the problem-solving onus into the tool’s geometry so that control becomes simpler. This motivates us to ask: can today’s foundation models offer useful priors to automatically invent—and effectively wield—such tools? We present VLMgineer, the first fully automatic framework designs tools and actions from scratch by harnessing the creativity of Vision–Language Models (VLMs) together with evolutionary search. We evaluate VLMgineer on a diverse benchmark of everyday manipulation scenarios that demand creative tool design and use. Across this suite, VLMgineer consistently discovers tools and policies that solve tasks more effectively and innovatively, transforming challenging robotics problems into straightforward executions. It also consistently outperforms VLM-generated designs from human specifications and existing human-crafted tools for everyday tasks. We further demonstrate that VLMgineer’s automatically designed tools and action policies transfer seamlessly to real-world task execution on a physical robot. To facilitate future research on automated tool invention, we will release our benchmark and code. Project Website: https://vlmgineer.github.io/.

Recently, considerable progress has been made in all-in-one image restoration. Generally, existing methods can be degradation-agnostic or degradation-aware. However, the former are limited in leveraging degradation estimation-based priors, and the latter suffer from the inevitable error in degradation estimation. Consequently, the performance of existing methods has a large gap compared to specific single-task models. In this work, we make a step forward in this topic, and present our UniRestorer with improved restoration performance. Specifically, we perform hierarchical clustering on degradation space, and train a multi-granularity mixture-of-experts (MoE) restoration model. Then, UniRestorer adopts both degradation and granularity estimation to adaptively select an appropriate expert for image restoration. In contrast to existing degradation-agnostic and -aware methods, UniRestorer can leverage degradation estimation to benefit degradation-specific restoration, and use granularity estimation to make the model robust to degradation estimation error. Experimental results show that our UniRestorer outperforms state-of-the-art all-in-one methods by a large margin, and is promising in closing the performance gap to specific single-task models. The code and pre-trained models will be publicly available.

The performance of fine-tuned language models is heavily influenced by the quality and quantity of their fine-tuning data. While scaling laws suggest that larger models benefit from more data during pretraining, the Less-is-More hypothesis highlights that downstream fine-tuning often requires only a small but high-quality dataset to effectively elicit a model’s pretrained knowledge. However, identifying such premium data, particularly in terms of difficulty and diversity, typically relies on human expertise, and existing methods offer limited guidance for automatic selection from large unannotated corpora. This work presents a novel quantitative framework that formalizes the interplay between question difficulty and diversity, and introduces *Difficulty–Diversity Collaborative Filtering* (DDCF): an automated approach that tailors data selection to the unique characteristics of each language model via collaborative filtering. By leveraging a small seed dataset to predict correctness across a large unannotated corpus, our method reduces the annotation cost by $100-200\times$, while maintaining downstream performance comparable to full-corpus fine-tuning.

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet optimally integrating Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through a comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from an entropy-based perspective, we reveal key differences between these paradigms: SFT induces coarse-grained, global shifts to policy distributions, while RL performs fine-grained, selective optimizations. Our analysis further establishes entropy as a critical indicator of training efficacy. Building on these observations, we introduce **S**upervised **R**einforcement **F**ine-**T**uning (**SRFT**), a single-stage framework that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. SRFT simultaneously applies SFT and RL to directly optimize LLMs using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT outperforms zero-RL baselines by **9.0%** on five mathematical reasoning benchmarks and by **10.9%** on three out-of-distribution benchmarks. Moreover, by leveraging demonstration data, SRFT maintains a more stable policy entropy, facilitating sustained policy improvement.

Noise is ubiquitous in data used to train large language models, but it is not well understood whether these models are able to correctly generalize to inputs generated without noise. Here, we study noise-robust learning: are transformers trained on data with noisy features able to find a target function that correctly predicts labels for noiseless features? We show that transformers succeed at noise-robust learning for a selection of $k$-sparse parity and majority functions, compared to LSTMs which fail at this task for even modest feature noise. However, we find that transformers typically fail at noise-robust learning of random $k$-juntas, especially when the boolean sensitivity of the optimal solution is smaller than that of the target function. We argue that this failure is due to a combination of two factors: transformers' bias toward simpler functions, combined with an observation that the empirically optimal function for noise-robust learning has lower sensitivity than the target function. We test this hypothesis by exploiting transformers' simplicity bias to trap them in an incorrect solution, but show that transformers can escape this trap by training with an additional loss term penalizing high-sensitivity solutions. Overall, we find that transformers are particularly ineffective for learning boolean functions in the presence of feature noise.

Federated Prompt Learning has emerged as a communication-efficient and privacy-preserving paradigm for adapting large vision-language models like CLIP across decentralized clients. However, the security implications of this setup remain underexplored. In this work, we present the first study of backdoor attacks in Federated Prompt Learning. We show that when malicious clients inject visually imperceptible, learnable noise triggers into input images, the global prompt learner becomes vulnerable to targeted misclassification while still maintaining high accuracy on clean inputs. Motivated by this vulnerability, we propose SABRE-FL, a lightweight, modular defense that filters poisoned prompt updates using an embedding-space anomaly detector trained offline on out-of-distribution data. SABRE-FL requires no access to raw client data or labels and generalizes across diverse datasets. We show, both theoretically and empirically, that malicious clients can be reliably identified and filtered using an embedding-based detector. Across five diverse datasets and four baseline defenses, SABRE-FL outperforms all baselines by significantly reducing backdoor accuracy while preserving clean accuracy, demonstrating strong empirical performance and underscoring the need for robust prompt learning in future federated systems.

Instruction-guided video editing has emerged as a rapidly advancing research direction, offering new opportunities for intuitive content transformation while also posing significant challenges for systematic evaluation. Existing video editing benchmarks fail to support the evaluation of instruction-guided video editing adequately and further suffer from limited source diversity, narrow task coverage and incomplete evaluation metrics. To address above limitations, we introduce IVEBench, a modern benchmark suite specifically designed for instruction-guided video editing assessment. IVEBench comprises a diverse database of 600 high-quality source videos, spanning seven semantic dimensions, and covering video lengths ranging from 32 to 1,024 frames. It further includes 8 categories of editing tasks with 35 subcategories, whose prompts are generated and refined through large language models and expert review. Crucially, IVEBench establishes a three-dimensional evaluation protocol encompassing video quality, instruction compliance and video fidelity, integrating both traditional metrics and multimodal large language model-based assessments. Extensive experiments demonstrate the effectiveness of IVEBench in benchmarking state-of-the-art instruction-guided video editing methods, showing its ability to provide comprehensive and human-aligned evaluation outcomes. All data and code will be made publicly available.

Interactive segmentation uses real-time user inputs, such as mouse clicks, to iteratively refine model predictions. Although not originally designed to address distribution shifts, this paradigm naturally lends itself to such challenges. In medical imaging, where distribution shifts are common, interactive methods can use user inputs to guide models towards improved predictions. Moreover, once a model is deployed, user corrections can be used to adapt the network parameters to the new data distribution, mitigating distribution shift. Based on these insights, we aim to develop a practical, effective method for improving the adaptive capabilities of interactive segmentation models to new data distributions in medical imaging. Firstly, we found that strengthening the model's responsiveness to clicks is important for the initial training process. Moreover, we show that by treating the post-interaction user-refined model output as pseudo-ground-truth, we can design a lean, practical online adaptation method that enables a model to learn effectively across sequential test images. The framework includes two components: (i) a Post-Interaction adaptation process, updating the model after the user has completed interactive refinement of an image, and (ii) a Mid-Interaction adaptation process, updating incrementally after each click. Both processes include a Click-Centered Gaussian loss that strengthens the model's reaction to clicks and enhances focus on user-guided, clinically relevant regions. Experiments on 5 fundus and 4 brain‑MRI databases show that our approach consistently outperforms existing methods under diverse distribution shifts, including unseen imaging modalities and pathologies. Code and pretrained models will be released upon publication.

Articulated objects are ubiquitous in daily life. Our goal is to achieve a high-quality reconstruction, segmentation of independent moving parts, and analysis of articulation. Recent methods analyse two different articulation states and perform per-point part segmentation, optimising per-part articulation using cross-state correspondences, given a priori knowledge of the number of parts. Such assumptions greatly limit their applications and performance. Their robustness is reduced when objects cannot be clearly visible in both states. To address these issues, in this paper, we present a new framework, *Articulation in Motion (AiM)*. We infer part-level decomposition, articulation kinematics, and reconstruct an interactive 3D digital replica from a user–object interaction video and a start-state scan. We propose a dual-Gaussian scene representation that is learned from an initial 3DGS scan of the object and a video that shows the movement of separate parts. It uses motion cues to segment the object into parts and assign articulation joints. Subsequently, a robust, sequential RANSAC is employed to achieve part mobility analysis \textit{without any part-level structural priors}, which clusters moving primitives into rigid parts and estimates kinematics while automatically determining the number of parts. The proposed approach separates the object into parts, each represented as a 3D Gaussian set, enabling high-quality rendering. Our approach yields higher quality part segmentation than previous methods, without prior knowledge. Extensive experimental analysis on both simple and complex objects validates the effectiveness and strong generalisation ability of our approach. Project page: https://haoai-1997.github.io/AiM/.

Electroencephalography (EEG) provides real-time insights into brain activity and supports diverse applications in neuroscience. While EEG foundation models (EFMs) have emerged to address the scalability issues of task-specific models, current approaches still yield clinically uninterpretable and weakly discriminative representations, inefficiently capture global dependencies, and neglect important local neural events. We present CodeBrain, a two-stage EFM designed to fill this gap. In the first stage, we introduce the TFDual-Tokenizer, which decouples heterogeneous temporal and frequency EEG signals into discrete tokens, quadratically expanding the representation space to enhance discriminative power and offering domain-specific representation-level interpretability by suggesting potential links to neural events and spectral rhythms. In the second stage, we propose the multi-scale EEGSSM architecture, which combines structured global convolution with sliding window attention to efficiently capture both sparse long-range and local dependencies, reflecting the brain’s small-world topology. Pretrained on the largest public EEG corpus, CodeBrain achieves strong generalization across 8 downstream tasks and 10 datasets under distribution shifts, supported by comprehensive ablations, scaling-law analyses, and interpretability evaluations.