ICLR 2026 Papers

Recent multimodal large language models (MLLMs) increasingly integrate multiple vision encoders to improve performance on various benchmarks, assuming that diverse pretraining objectives yield complementary visual signals. However, we show this assumption often fails in practice. Through systematic encoder masking across representative multi-encoder MLLMs, we find that performance typically degrades gracefully—and sometimes even improves—when selected encoders are masked, revealing pervasive encoder redundancy. To quantify this effect, we introduce two principled metrics: the Conditional Utilization Rate (CUR), which measures an encoder’s marginal contribution in the presence of others, and the Information Gap (IG), which captures heterogeneity in encoder utility within a model. Using these tools, we observe: (i) strong specialization on tasks like OCR \& Chart, where a single encoder can dominate with a CUR >90%, (ii) high redundancy on general VQA and knowledge-based tasks, where encoders are largely interchangeable, (iii) instances of detrimental encoders with negative CUR. Notably, masking specific encoders can yield up to 16% higher accuracy on a specific task category and 3.6% overall performance boost compared to the full model. Furthermore, single- and dual- encoder variants recover over 90% of baseline on most non-OCR tasks. Our analysis challenges the “more encoders are better” heuristic in MLLMs and provides actionable diagnostics for developing more efficient and effective multimodal architectures.

Diffusion Policy has demonstrated strong visuomotor modeling capabilities, but its high computational cost renders it impractical for real-time robotic control. Despite huge redundancy across repetitive denoising steps, existing diffusion acceleration techniques fail to generalize to Diffusion Policy due to fundamental architectural and data divergences. In this paper, we propose **B**lock-wise **A**daptive **C**aching (**BAC**), a method to accelerate Diffusion Policy by caching intermediate action features. BAC achieves lossless action generation acceleration by adaptively updating and reusing cached features at the block level, based on a key observation that feature similarities exhibit non-uniform temporal dynamics and distinct block-specific patterns. To operationalize this insight, we first design an Adaptive Caching Scheduler to identify optimal update timesteps by maximizing the global feature similarities between cached and skipped features. However, applying this scheduler for each block leads to significant error surges due to the inter-block propagation of caching errors, particularly within Feed-Forward Network (FFN) blocks. To mitigate this issue, we develop the Bubbling Union Algorithm, which truncates these errors by updating the upstream blocks with significant caching errors before downstream FFNs. As a training-free plugin, BAC is readily integrable with existing transformer-based Diffusion Policy and vision-language-action models. Extensive experiments on multiple robotic benchmarks demonstrate that BAC achieves up to 3x inference speedup for free. Project page: https://block-wise-adaptive-caching.github.io.

Humans often rely on visual aids, such as diagrams or sketches, when tackling complex problems. Teaching multimodal models to adopt similar strategies, a process known as Visual Chain of Thought (visual CoT), is much more difficult. The main challenges are: (1) weak performance of off-the-shelf visual CoT, which hinders reinforcement learning, and (2) the lack of high-quality visual CoT training data. We introduce **Zebra-CoT** a diverse large-scale interleaved text-image reasoning dataset with 182,384 reasoning traces across 18 domains with over 50 distinct tasks. This dataset is specifically designed to train models to natively perform visual CoT. We emphasize four categories of tasks where sketching or visual reasoning is especially natural, spanning (a) *scientific questions* such as geometry, physics, and algorithms; (b) *2D visual reasoning tasks* like visual search and jigsaw puzzles; (c) *3D reasoning tasks* including 3D multi-hop inference, embodied and robot planning; and (d) *visual logic problems and strategic games* like chess. Fine-tuning Anole‑7B model on Zebra-CoT yields a +12\% improvement in our test‑set accuracy and up to +13\% performance gains on standard VLM benchmarks. Similarly, fine-tuning Bagel‑7B produces models capable of generating high-quality interleaved visual reasoning chains, underscoring Zebra-CoT's effectiveness in advancing multimodal reasoning.

Recent advance in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), one state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance boost while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA forces a loop order that is only efficient with a relatively large number of query heads in each Grouped Query Attention (GQA) group, whereas existing LLMs widely adopt much smaller number of query heads in each GQA group --- such an inconsistency significantly limits the applicability of this sparse algorithmic advance. In this work, we propose **F**lash **S**parse **A**ttention (**FSA**), an alternative kernel implementation that enables efficient NSA computation across a wide range of popular LLMs with varied smaller number of query heads in each GQA group on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5x and on average 1.6x kernel-level latency reduction, (ii) up to 1.25x and 1.09x on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36x and 1.11x on average for prefill-phase speedup in LLM generative inference.

Attention sinks and compression valleys have attracted significant attention as two puzzling phenomena in large language models, but have been studied in isolation. In this work, we present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream. We prove theoretically that massive activations necessarily produce representational compression and establish bounds on the resulting entropy reduction. Through experiments across several models (410M--120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys and attention sinks emerge simultaneously. Targeted ablation validates our theoretical predictions. This unified view motivates us to propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations. Specifically, we posit that Transformer-based LLMs process tokens in three distinct phases: (1) broad mixing in the early layers, (2) compressed computation with limited mixing in the middle layers, and (3) selective refinement in the late layers. Our framework helps explain why embedding tasks perform best at intermediate layers, whereas generation tasks benefit from full-depth processing, clarifying differences in task-dependent representations.

As large language models continue to scale, their growing computational and storage demands pose significant challenges for real-world deployment. In this work, we investigate redundancy within Transformer-based models and propose an entropy-based pruning strategy to enhance efficiency while maintaining performance. Empirical analysis reveals that the entropy of hidden representations decreases in the early blocks but progressively increases across most subsequent blocks. This trend suggests that entropy serves as a more effective measure of information richness within computation blocks. Unlike cosine similarity, which primarily captures geometric relationships, entropy directly quantifies uncertainty and information content, making it a more reliable criterion for pruning. Extensive experiments demonstrate that our entropy-based pruning approach surpasses cosine similarity-based methods in reducing model size while preserving accuracy, offering a promising direction for efficient model deployment.

Fine-tuning Large Language Models (LLMs) with downstream data is often considered time-consuming and expensive. Structured pruning methods are primarily employed to improve the inference efficiency of pre-trained models. Meanwhile, they often require additional time and memory for training, knowledge distillation, structure search, and other strategies, making efficient model fine-tuning challenging to achieve. To simultaneously enhance the training and inference efficiency of downstream task fine-tuning, we introduce GradPruner, which can prune layers of LLMs guided by gradients in the early stages of fine-tuning. GradPruner uses the cumulative gradients of each parameter during the initial phase of fine-tuning to compute the Initial Gradient Information Accumulation Matrix (IGIA-Matrix) to assess the importance of layers and perform pruning. We sparsify the pruned layers based on the IGIA-Matrix and merge them with the remaining layers. Only elements with the same sign are merged to reduce interference from sign variations. We conducted extensive experiments on two LLMs across eight well-known datasets in downstream tasks. Including medical, financial, and general benchmark tasks. The results demonstrate that GradPruner has achieved a parameter reduction of 40% with only a 0.99% decrease in accuracy. Our code is available at https://anonymous.4open.science/r/LLM-GradPrune-436D.

High dynamic range (HDR) video can be reconstructed from low dynamic range (LDR) sequences with alternating exposures. However, most existing methods overlook the degradations (e.g., noise and blur) in LDR frames, focusing only on the brightness and position differences between them. To address this gap, we propose DeAltHDR, a novel framework for high-quality HDR video reconstruction from degraded sequences. Our framework addresses two key challenges. First, noisy and blurry contents complicate inter-frame alignment. To tackle this, we propose a flow-guided masked attention that leverages optical flow for a dynamic sparse cross-attention computation, achieving superior performance while maintaining efficiency. Notably, its controllable attention ratio allows for adaptive inference costs. Second, the lack of real-world paired data hinders practical deployment. We overcome this with a two-stage training paradigm: the model is first pre-trained on our newly introduced synthetic paired dataset and subsequently fine-tuned on unlabeled real-world videos via a proposed self-supervised method. Experiments show our method outperforms state-of-the-art ones. The datasets and code will be publicly available.

Machine unlearning (MU) for large language models (LLMs), commonly referred to as LLM unlearning, seeks to remove specific undesirable data or knowledge from a trained model, while maintaining its performance on standard tasks. While unlearning plays a vital role in protecting data privacy, enforcing copyright, and mitigating sociotechnical harms in LLMs, we identify a new vulnerability post-unlearning: unlearning trace detection. We discover that unlearning leaves behind persistent "fingerprints" in LLMs, detectable traces in both model behavior and internal representations. These traces can be identified from output responses, even when prompted with forget-irrelevant inputs. Specifically, even a simple supervised classifier can determine whether a model has undergone unlearning, using only its prediction logits or even its textual outputs. Further analysis shows that these traces are embedded in intermediate activations and propagate nonlinearly to the final layer, forming low-dimensional, learnable manifolds in activation space. Through extensive experiments, we demonstrate that unlearning traces can be detected with over 90% accuracy even under forget-irrelevant inputs, and that larger LLMs exhibit stronger detectability. These findings reveal that unlearning leaves measurable signatures, introducing a new risk of reverse-engineering forgotten information when a model is identified as unlearned, given an input query.

As the parameter size of large language models (LLMs) continues to grow, the latency of autoregressive inference increases due to memory-bound computational inefficiency. To address this, speculative decoding has been proposed, where a large target model verifies multiple tokens generated in parallel by a smaller draft model. However, the performance of speculative decoding is fundamentally limited by the draft model’s capacity, which stems from the parameter gap between the two models. To overcome this limitation, we propose RepSpec, which combines structural re-parameterization with draft model training. During training, redundant linear structures are introduced and later merged into the backbone network during inference, thus enhancing the draft model’s training effectiveness without increasing inference cost. By applying our method to improve the current state-of-the-art approach, EAGLE, we achieve a significant improvement in accepted sequence length. Furthermore, considering the specific characteristics of the speculative decoding scenario, we explore a hybrid training strategy that combines linear and nonlinear structures, which yields a further improvement in acceptance length.

Tool planning with large language models (LLMs), referring to selecting, organizing, and preparing the tools necessary to complete a user request, bridges the gap between natural language understanding and task execution. However, current works treat different tools as isolated components and fail to leverage the inherent dependencies of tools, leading to invalid planning results. Since tool dependencies are often incomplete, it becomes challenging for LLMs to accurately identify the appropriate tools required by a user request, especially when confronted with a large toolset. To solve this challenge, we propose GTool, which is the first work aiming to enhance the tool planning ability of LLMs under incomplete dependencies. GTool constructs a request-specific tool graph to select tools efficiently and generate the \<graph token\> which provides sufficient dependency information understandable by LLMs. Moreover, a missing dependency prediction task is designed to improve the reliability of GTool with incomplete dependencies. Without trimming LLMs, GTool can be seamlessly integrated with various LLM backbones without extensive retraining. Extensive experiments show that GTool achieves more than 29.6% performance improvements compared with the state-of-the-art (SOTA) baselines with a light-weight (7B) LLM backbone.

Accurate tumor analysis is central to clinical radiology and precision oncology, where early detection, reliable lesion characterization, and pathology-level risk assessment directly guide diagnosis, staging, and treatment planning. Chain-of-Thought (CoT) reasoning is particularly critical in this setting, as it enables stepwise interpretation from imaging findings to clinical impressions and pathology-level conclusions, ensuring traceability and reducing diagnostic errors. Here, we target the clinical tumor analysis task and build a large-scale benchmark that operationalizes a multimodal reasoning pipeline, spanning findings, impressions, and pathology predictions. We curate TumorCoT, a large-scale dataset of 1.5M CoT-labeled VQA instructions paired with 3D CT scans, with step-aligned rationales and cross-modal alignments along the “findings → impression → pathology” trajectory, enabling standardized evaluation of both final accuracy and reasoning consistency. We further propose TumorChain, a multimodal interleaved reasoning framework that tightly couples 3D imaging encoders, clinical text understanding, and organ-level vision-language alignment. Through cross-modal alignment and iterative interleaved causal reasoning, TumorChain grounds visual evidence, aggregates conclusions, and issues pathology predictions after multiple rounds of self-refinement, improving traceability and reducing hallucination risk. TumorChain demonstrates consistent gains over strong unimodal and pipeline baselines in lesion detection, impression quality, and pathology classification, and successfully generalizes to the public DeepTumorVQA benchmark. Ablations validate the key contributions of interleaved reasoning and clinical CoT. Clinically, these advances lay the groundwork for reliable, interpretable tumor assessment to support real-world decision-making. We release the task, benchmark, and evaluation protocol to advance safe, explainable, and reproducible multimodal reasoning for high-stakes tumor analysis. Our project is available at https://anonymous.4open.science/r/TumorChain-D6E6.

Advances in diffusion, autoregressive, and hybrid models have enabled high-quality image synthesis for tasks such as text-to-image, editing, and reference-guided composition. Yet, existing benchmarks remain limited, either focus on isolated tasks, cover only narrow domains, or provide opaque scores without explaining failure modes. We introduce \textbf{ImagenWorld}, a benchmark of 3.6K condition sets spanning six core tasks (generation and editing, with single or multiple references) and six topical domains (artworks, photorealistic images, information graphics, textual graphics, computer graphics, and screenshots). The benchmark is supported by 20K fine-grained human annotations and an explainable evaluation schema that tags localized object-level and segment-level errors, complementing automated VLM-based metrics. Our large-scale evaluation of 14 models yields several insights: (1) models typically struggle more in editing tasks than in generation tasks, especially in local edits. (2) models excel in artistic and photorealistic settings but struggle with symbolic and text-heavy domains such as screenshots and information graphics. (3) closed-source systems lead overall, while targeted data curation (e.g., Qwen-Image) narrows the gap in text-heavy cases. (4) modern VLM-based metrics achieve Kendall accuracies up to 0.79, approximating human ranking, but fall short of fine-grained, explainable error attribution. ImagenWorld provides both a rigorous benchmark and a diagnostic tool to advance robust image generation.

Recent advances in foundation models highlight a clear trend toward unification and scaling, showing emergent capabilities across diverse domains. While image generation and editing have rapidly transitioned from task-specific to unified frameworks, video generation and editing remain fragmented due to architectural limitations and data scarcity. In this work, we introduce EditVerse, a unified framework for image and video generation and editing within a single model. By representing all modalities, i.e., text, image, and video, as a unified token sequence, EditVerse leverages self-attention to achieve robust in-context learning, natural cross-modal knowledge transfer, and flexible handling of inputs and outputs with arbitrary resolutions and durations. To address the lack of video editing training data, we design a scalable data pipeline that curates 232K video editing samples and combines them with large-scale image and video datasets for joint training. Furthermore, we present EditVerseBench, the first benchmark for instruction-based video editing covering diverse tasks and resolutions. Extensive experiments and user studies demonstrate that EditVerse achieves state-of-the-art performance, surpassing existing open-source and commercial models, while exhibiting emergent editing and generation abilities across modalities.

Recent advances in multimodal learning have achieved remarkable success across diverse vision–language tasks. However, such progress heavily relies on large-scale image–text datasets, making training costly and inefficient. Prior efforts in dataset filtering and pruning attempt to mitigate this issue, but still require relatively large subsets to maintain performance and fail under very small subsets. Dataset distillation offers a promising alternative, yet existing multimodal dataset distillation methods require full-dataset training and joint optimization of image pixels and text features, making them architecture-dependent and limiting cross-architecture generalization. To overcome this, we propose a learning-free dataset distillation framework that eliminates the need for large-scale training and optimization while enhancing generalization across architectures. Our method uses CLIP to extract aligned image–text embeddings, obtains prototypes, and employs an unCLIP decoder to synthesize images, enabling efficient and scalable multimodal dataset distillation. Extensive experiments demonstrate that our approach consistently outperforms optimization-based dataset distillation and subset selection methods, achieving state-of-the-art cross-architecture generalization.

We introduce SCUBA, a benchmark designed to evaluate computer-use agents on customer relationship management (CRM) workflows within the Salesforce platform. SCUBA contains 300 task instances derived from real user interviews, spanning three primary personas—platform administrators, sales representatives, and service agents. The tasks test a range of enterprise-critical abilities, including Enterprise Software UI navigation, data manipulation, workflow automation, information retrieval, and troubleshooting. To ensure realism, SCUBA operates in Salesforce sandbox environments with support for parallel execution and fine-grained evaluation metrics to capture milestone progress. We benchmark a diverse set of agents under both zero-shot and demonstration-augmented settings. We observed huge performance gaps in different agent design paradigm and gaps between the open-source model and the closed-source model. In the zero-shot setting, open-source model powered computer-use agents that have strong performance on related benchmarks like OSWorld only have less than 5\% success rate on SCUBA, while methods built on closed-source models can still have up to 39\% percent task success rate. In the demonstration-augmented settings, task success rates can be improved to 50\% while simultaneously reducing time and costs by 13\% and 16\%, respectively. These findings highlight both the challenges of enterprise tasks automation and the promise of agentic solutions. By offering a realistic benchmark with interpretable evaluation, SCUBA aims to accelerate progress in building reliable computer-use agents for complex business software ecosystems.

Multimodal Large Language Models (MLLMs) hold promise for accelerating scientific discovery by interpreting complex experimental procedures. However, their true capabilities are poorly understood, as existing benchmarks neglect the fine-grained and long-horizon nature of authentic laboratory work, especially in wet-lab settings. To bridge this gap, we introduce ExpVid, the first benchmark designed to systematically evaluate MLLMs on scientific experiment videos. Curated from peer-reviewed video publications, ExpVid features a new three-level task hierarchy that mirrors the scientific process: (1) Fine-grained Perception of tools, materials, and actions; (2) Procedural Understanding of step order and completeness; and (3) Scientific Reasoning that connects the full experiment to its published conclusions. Our vision-centric annotation pipeline, combining automated generation with multi-disciplinary expert validation, ensures that tasks require visual grounding. We evaluate 19 leading MLLMs on ExpVid and find that while they excel at coarse-grained recognition, they struggle with disambiguating fine details, tracking state changes over time, and linking experimental procedures to scientific outcomes. Our results reveal a notable performance gap between proprietary and open-source models, particularly in high-order reasoning. ExpVid not only provides a diagnostic tool but also charts a roadmap for developing MLLMs capable of becoming trustworthy partners in scientific experimentation.

Historically, neuroscience has progressed by fragmenting into specialized domains, each focusing on isolated modalities, tasks, or brain regions. While fruitful, this approach hinders the development of a unified model of cognition. Here, we introduce TRIBE, the first deep neural network trained to predict brain responses to stimuli across multiple modalities, cortical areas and individuals. By combining the pretrained representations of text, audio and video foundational models and handling their time-evolving nature with a transformer, our model can precisely model the spatial and temporal fMRI responses to videos, achieving the first place in the Algonauts 2025 brain encoding competition with a significant margin over competitors. Ablations show that while unimodal models can reliably predict their corresponding cortical networks (e.g. visual or auditory networks), they are systematically outperformed by our multimodal model in high-level associative cortices. Currently applied to perception and comprehension, our approach paves the way towards building an integrative model of representations in the human brain. Our code is available at \url{https://anonymous.4open.science/r/algonauts-2025-C63E}.

Directed Acyclic Graphs (DAGs) are a standard tool in causal modeling, but their suitability for capturing the complexity of large-scale multimodal data is questionable. In practice, real-world multimodal datasets are often collected from heterogeneous generative processes that do not conform to a single DAG. Instead, they may involve multiple, and even opposing, DAG structures with inverse causal directions. To address this gap, in this work, we first propose a novel latent partial causal model tailored for multimodal data representation learning, featuring two latent coupled variables parts connected by an undirected edge, to represent the transfer of knowledge across modalities. Under specific statistical assumptions, we establish an identifiability result, demonstrating that representations learned by MultiModal Contrastive Learning (MMCL) correspond to the latent coupled variables up to a trivial transformation. This result deepens our understanding of the why MMCL works, highlights its potential for representation disentanglement, and expands the utility of pre-trained models like CLIP. Synthetic experiments confirm the robustness of our findings, even when the assumptions are partially violated. Most importantly, experiments on a pre-trained CLIP model embodies disentangled representations, enabling few-shot learning and improving domain generalization across diverse real-world datasets. Together, these contributions push the boundaries of MMCL, both in theory and in practical applications.