ICML 2026 Papers

Visual perception in modern Vision-Language Models (VLMs) is constrained by a perceptual bandwidth bottleneck: a broad field of view preserves global context but sacrifices the fine-grained details required for complex reasoning. We argue that high-resolution visual reasoning is therefore not only semantic reasoning but also task-relevant evidence acquisition under limited perceptual bandwidth. Inspired by active vision and information foraging, we formalise this process as sequential Bayesian optimal experimental design (S-BOED), where an agent decides which visual evidence to acquire before answering. Since exact Bayesian inference is intractable in continuous gigapixel spaces, we derive a tractable coverage-resolution objective as a proxy for task-relevant information gain. We instantiate this framework with FOVEA, a training-free procedure that refines VLM crop proposals through evidence-oriented probing. Experiments on high-resolution benchmarks show consistent gains over direct and ReAct-style baselines, with particularly strong improvements in search-dominated remote-sensing settings.

The rapid rise of generative AI has made multimodal fake news increasingly realistic and pervasive, posing severe threats to public trust and social stability. Existing detection methods rely heavily on manipulation-specific models and large-scale labeled data, resulting in poor generalization to emerging manipulation types. We observed that the essence of manipulated misinformation lies in its intrinsic conflicts, i.e., semantic or physical inconsistencies either across modalities or with common world knowledge. Inspired by this observation, we propose Conflict-{Oriented REasoning (CORE) framework, an effective paradigm that learns to endows multimodal large language models (MLLMs) with explicit conflict-capturing capability. To this end, CORE first constructs the Conflict Attribution Corpus (CAC) with fine-grained annotations of conflict factors and sources, providing essential data support for subsequent conflict perception training. By performing conflict-oriented representation enhancement and reasoning based on CAC, CORE achieves robust and generalizable conflict detection, effectively and rapidly adapting to unseen manipulation types with a few samples or in even zero-shot settings. Extensive experiments demonstrate that CORE surpasses state-of-the-art models by 9.7\%, 14.1\%, and 11.8\% in accuracy on the DGM$^4$, MMFakeBench, and MDSM benchmarks, respectively.

Activation sparsity offers a compelling route to accelerate large language model (LLM) inference by selectively suppressing hidden activations, yet existing approaches exhibit severe accuracy degradation at high sparsity. We show that this failure stems from representational instability: *activation sparsity disrupts input-dependent activation learned during pretraining, inducing distribution shifts in hidden states*. We address this issue by reframing activation sparsity as a representational alignment problem and introducing **Spontaneous Neurons (SPON)**, a lightweight mechanism inspired by spontaneous neural activity in biological systems. SPON injects a small set of learnable, input-independent activation vectors that act as persistent representational anchors for sparse computation. These vectors are trained via distribution matching to the dense model and can be absorbed into bias terms after training, incurring negligible inference overhead. Across multiple LLM backbones, SPON consistently restores performance, stabilizes latent representations, and preserves generalization. Our results establish SPON as an effective and principled solution for reliable activation-sparse inference, and offer new insights into knowledge retention in LLMs.

Multi-model learning has attracted great attention in visual-text tasks. However, visual-tabular data, which plays a pivotal role in high-stakes domains like healthcare and industry, remains underexplored. In this paper, we introduce \textit{VT-Bench}, the first unified benchmark for standardizing vision-tabular discriminative prediction and generative reasoning tasks. VT-Bench aggregates 14 datasets across 9 domains (medical-centric, while covering pets, media, and transportation) with over 756K samples. We evaluate 23 representative models, including unimodal experts, specialized visual-tabular models, general-purpose vision-language models (VLMs), and tool-augmented methods, highlighting substantial challenges of visual-tabular learning. We believe VT-Bench will stimulate the community to build more powerful multi-modal vision-tabular foundation models. Benchmark: \url{https://github.com/LAMDA-NeSy/VT-Bench}

SQL is central to enterprise data engineering, yet generating fully correct SQL code in a single attempt remains difficult—even for experienced developers and advanced Text-to-SQL LLMs—often requiring multiple debugging iterations. We introduce Squirrel Benchmark, the first benchmark for enterprise-level SQL reasoning and debugging. Our benchmark is built upon two key innovations: (1) an automated construction workflow that employs reverse engineering to systematically inject realistic bugs into large-scale SQL code, enabling scalable and diverse benchmark generation; and (2) an execution-free evaluation framework tailored for enterprise settings, providing fast, accurate, and resource-efficient assessment. Squirrel comprises 469 Squirrel-Syntax queries featuring syntax errors with explicit error messages, and 516 Squirrel-Semantic queries targeting semantic errors where code fails to meet user intent. The queries are highly complex, averaging over 140 lines and featuring deep, wide abstract syntax trees (average width >11, depth >8.7). Evaluation of nearly 30 LLMs reveals a substantial performance gap: the best-performing model, Claude-4-Sonnet, achieves only 36.46% accuracy on Squirrel-Syntax and 32.17% on Squirrel-Semantic, while most models score below 20%. We further explore four solution strategies, identify key challenges, and outline promising directions for enterprise SQL debugging with LLMs.

Knowledge-Based Visual Question Answering (KB-VQA) relies on retrieval from large-scale knowledge bases, yet this stage is often treated simplistically. Existing methods typically adopt single-view indexing or naive multi-view fusion, leading to systematic coverage gaps. In this work, we demonstrate that different views exhibit strong complementarity in retrieval. Motivated by this observation, we propose SCAR, a Structure-aware Cross-View Retrieval framework that exploits cross-view structural complementarity at inference time without additional training. SCAR enhances retrieval via structure-aware similarity propagation within each view and explicit cross-view redundancy regulation. Experiments on multiple KB-VQA benchmarks demonstrate that SCAR substantially improves retrieval recall, approaches retrieval coverage upper bounds, and consistently boosts end-to-end KB-VQA performance with negligible inference overhead.

We find that current emotion-oriented Omni-MLLMs still lack *reliable omni-modal perception*: they (i) underutilize multimodal cues in their reasoning trajectories and (ii) exhibit unfaithful behavior, often hallucinating modality-specific statements from other modalities. Building on these insights, we propose **OPPO** (**O**mni-**P**erception **P**olicy **O**ptimization), a reinforcement learning framework that explicitly optimizes multimodal perception. First, an Omni-Perception Reward decomposes ground-truth reasoning into fine-grained visual, acoustic, and emotion cues and rewards trajectories that semantically recover these cues. Second, an Omni-Perception Loss compares the policy under full and unimodally masked inputs, applying a KL penalty only to modality-specific evidence tokens to suppress cross-modal hallucination. We further introduce *MEP-Bench*, a diagnostic benchmark that quantifies *utilization* and *faithfulness*. Experiments show that OPPO achieves state-of-the-art performance on MER-UniBench and substantially improves utilization and faithfulness scores on MEP-Bench, highlighting the importance of sufficient and faithful omni perception for multimodal emotion reasoning.

Real-world agricultural counting often operates in the extreme regime of \textbf{Dense and Indiscernible Object Counting (DIOC)}, where targets are tiny, clustered, and highly camouflaged. To facilitate research in this domain, we introduce \textbf{DIOCblueberry}, a large-scale benchmark that pushes the boundaries of visual perception. Unlike general datasets with salient objects, DIOCblueberry features extreme occlusion and camouflage. Compared to the popular FSC147 benchmark, it contains \textbf{1.9$\times$ more instances} per image (avg. 108) with an average box pixel ratio that is \textbf{7.9$\times$ smaller}, serving as a rigorous testbed for model robustness. Standard counting methods struggle in these scenarios due to severe visual ambiguity and scale mismatch. To address this, we propose \textbf{MaskCount}, a coarse-to-fine framework that incorporates semantic guidance. MaskCount leverages Vision-Language Models (CLIP) to generate pseudo segmentation masks for background suppression and employs a contrastive loss to maximize feature discriminability between fruits and foliage. Additionally, we design an edge-aware cropping mechanism to resolve boundary truncation in dense clusters. Extensive experiments demonstrate that MaskCount achieves a new state-of-the-art, reducing MAE and RMSE by \textbf{49.16\%} and \textbf{70.50\%} respectively on DIOCblueberry, with strong generalization to other agricultural scenes.

Large vision-language models (LVLMs) have demonstrated remarkable capabilities across a wide range of multimodal reasoning tasks. However, recent research shows that they are susceptible to adversarial examples. Existing LVLM attack methods are generally deployed in the white- or black-box setting, which severely rely on full-model gradients or elaborated transfer strategies, resulting in large resource costs. To this end, this paper focuses on a more efficient gray-box attack setting by solely accessing LVLM's vision encoder. Instead of using target images as the adversarial guidance, our main goal is to perturb the visual feature to best match more natural attacker-chosen target texts. Specifically, we develop a global semantic alignment module to project the visual features onto the SVD-structured subspace spanned by the textual semantics. We also propose to align detailed visual features with multi-context semantic texts extended by LLMs over discrete distributions via optimal transport. Extensive experiments demonstrate the superiority of the proposed method, while our attack is further proven to achieve great transferability across various LVLMs with CLIP-aware transfer designs.

The visual modality of vision-language models (VLMs) is an underexplored attack surface for bypassing safety alignment. We introduce four jailbreak attacks exploiting the vision component: (1) encoding harmful instructions as visual symbol sequences with a decoding legend, (2) replacing harmful objects with benign substitutes (e.g., bomb $\rightarrow$ banana) then prompting for harmful actions using the substitute term, (3) replacing harmful text in images (e.g., on book covers) with benign words while visual context preserves the original meaning, and (4) visual analogy puzzles whose solution requires inferring a prohibited concept. Evaluating across six frontier VLMs, our visual attacks bypass safety alignment and expose a cross-modality alignment gap: text-based safety training does not automatically generalize to harmful intent conveyed visually. For example, our visual cipher achieves 40.9\% attack success on Claude-Haiku-4.5 versus 10.7\% for an equivalent textual cipher. To further our insight into the attack mechanism, we present preliminary interpretability and mitigation results. These findings highlight that robust VLM alignment requires treating vision as a first-class target for safety post-training.

Online model editing for multimodal large language models (MLLMs) requires assimilating a stream of corrections under tight compute and memory budgets. Yet editors developed for text-only LLMs often degrade on MLLMs: visually dominant activations skew the statistics that shape updates, causing *cross-modal conflict*, while sequential writes become entangled in a shared edit space and amplify long-horizon interference, causing *inter-edit interference*. To address these, we propose **M-ORE**, a modality-decoupled online recursive editor for lifelong MLLM adaptation. M-ORE is derived from a unified proximal-projection formulation and admits a closed-form update with a Sherman-Morrison recursion, yielding constant per-edit overhead. It maintains module-wise locality statistics for the text stack and the visual projector to avoid visually dominated update shaping and performs continual updates in a fixed orthogonal low-rank edit subspace via a Sherman-Morrison recursion to mitigate long-horizon interference. Experiments on multiple MLLM backbones and online editing benchmarks show that our M-ORE method consistently improves reliability, generality, and locality over strong baselines, while achieving favorable quality-efficiency scaling. Our code is publicly available at https://github.com/lab-klc/M-ORE.

Inference-time scaling has emerged as a critical avenue for enhancing Large Language Models' performance, yet real-world deployment is constrained by strict computational budgets. In this work, we formulate inference budget allocation as a global constrained optimization problem governed by economic principles. By modeling per-query reasoning utility with a shifted-surge function, we derive an optimal allocation policy based on a global shadow price that equilibrates marginal utility under resource scarcity. Based on this theory, we propose Constrained Latent-utility Equilibrium Allocation for Reasoning (CLEAR). It performs rational abandonment and reallocates resources from insolvent queries to solvable queries near their emergence thresholds. Extensive experiments on several reasoning tasks with different traffic streams demonstrate that CLEAR significantly improves the Pareto frontier of total token cost versus mean accuracy. In resource-scarce regimes, CLEAR achieves up to a 3× improvement in global accuracy compared to uniform allocation.

Representations of the world, arguably, contain information about features (e.g. something is blue, something is a circle) but also information about which features are part of the same object (e.g. the circle is blue), which we call binding information. Any system with the ability to understand scenes with multiple objects must be able to solve the binding problem: it needs to know which features belong together. However, despite work showing that Vision Transformers (ViTs) know which patches belong together, it is not known whether current deep learning models learn to exhibit binding information, i.e., for features. We may believe that there is not much binding information, after all misattributing features to wrong objects is a common failure of ViT-based architectures, especially in scenes with objects sharing features. Here we formalize the binding problem with an information-theoretic approach, and introduce a probing method to measure binding information in model representations. We perform experiments on ViTs, measuring binding from different components of the architecture, such as the image summary token [CLS] or the spatial tokens. We use datasets with different binding challenges, such as feature sharing, occlusion, and natural features, while comparing the performance of several pre-trained ViTs. Overall, our research demonstrates binding as a key ingredient to strong visual recognition and reasoning.

While AI-generated hallucinations pose considerable risks, the underlying cognitive mechanisms by which humans can successfully recognize or be misled by these hallucinations remain unclear. To address this problem, this paper explores humans' neural dynamics to characterize how the brain processes hallucinated content. We record EEG signals from 27 participants while they are performing a verification task to judge the correctness of image descriptions generated by a multi-modal large language model (MLLM). Based on an averaged event-related potential (ERP) study, we reveal that multiple cognitive processes, e.g., semantic integration, inferential processing, memory retrieval, and cognitive load, exhibit distinct patterns when humans process hallucinated versus non-hallucinated content. Notably, neural responses to hallucinations that were misjudged versus correctly judged by human participants showed significant differences. This indicates that misjudged AI-generated hallucinations failed to trigger the standard neurocognitive fact verification pathway. The detailed code can be accessed openly through https://github.com/Promise-Z5Q2SQ/EEG-Hallucination.

Diffusion transformers (DiTs) equipped with multimodal attention (MM-Attn) have become a dominant paradigm for image generation. However, preventing the generation of harmful content remains a critical challenge, particularly in image-to-image (I2I) editing tasks. Existing safety mechanisms are primarily designed for text-to-image (T2I) synthesis or U-Net-based architectures, which limits their effectiveness for unified safety mitigation in DiT-based frameworks. To bridge this gap, we propose Unified Visual Safety Regulator (UVR), a training-free safe generation framework that regulates unsafe semantics in generated images. UVR is grounded in an analysis of attention dynamics from the perspective of information flow in MM-Attn. We identify a task-independent start-up stage, during which unsafe semantics in output patches rapidly emerge and can be accurately localized, followed by task-specific semantic amplification and interference stages, where harmful signals are further propagated and entangled with benign content. Based on these observations, UVR mitigates unsafe generation through unified, targeted attention modulation and explicit restriction of harmful information flow over the identified unsafe output patches. Experiments across various concepts show that UVR achieves state-of-the-art safety performance by achieving 91% and 77% erase rate in image synthesis and editing tasks, while preserving visual quality and fidelity with minimal degradation. Code is available at https://github.com/deng12yx/UVR.

Bimanual dexterous manipulation relies on integrating multimodal inputs to perform complex real-world tasks. To address the challenges of effectively combining these modalities, we propose DECO, a decoupled multimodal diffusion transformer that disentangles vision, proprioception, and tactile signals through specialized conditioning pathways, enabling structured and controllable integration of multimodal inputs, with a lightweight adapter for parameter-efficient injection of additional signals. Alongside DECO, we release DECO-50 dataset for bimanual dexterous manipulation with tactile sensing, consisting of 50 hours of data and over 5M frames, collected via teleoperation on real dual-arm robots. We train DECO on DECO-50 and conduct extensive real-world evaluation with over 2,000 robot rollouts. Experimental results show that DECO achieves the best performance across all tasks, with a 72.25\% average success rate and a 21\% improvement over the baseline. Moreover, the tactile adapter brings an additional 10.25\% average success rate across all tasks and a 20\% gain on complex contact-rich tasks while tuning less than 10\% of the model parameters.

Code-as-Policy (CaP) is a paradigm in which a language or vision-language model generates executable robot control programs, yet its effectiveness as an autonomous controller for embodied manipulation remains underexplored. Prior CaP systems often rely on high-level, human-designed primitives, making it difficult to separate agent capability from designer-provided scaffolding. We present CaP-X, an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. CaP-X includes four components. CaP-Gym is an interactive environment in which coding agents control robots by synthesizing and executing programs that compose perception and control primitives. Building on this foundation, CaP-Bench evaluates frontier language and vision-language models across varying levels of abstraction, interaction, and perceptual grounding. Across 12 models, the task success rates improve with human-crafted abstractions but degrade as these priors are removed, exposing a dependence on designer scaffolding. At the same time, we observe that scaling test-time computation with multi-turn interaction, structured execution feedback, visual differencing, automatic skill synthesis, and ensembled reasoning can substantially improve robustness even when agents operate over low-level primitives. These findings motivate CaP-Agent0, a training-free framework that achieves near human-level reliability on several manipulation tasks in simulation and on real embodiments. CaP-RL explores reinforcement learning with verifiable rewards to improve success rates and supports sim-to-real transfer through a shared code-as-action-space interface. Together, CaP-X provides an open-access platform for advancing embodied coding agents. Project page: https://capgym.github.io

Multimodal large language models (MLLMs) have shown strong potential as embodied agents, yet embodied geo-localization remains underexplored due to the lack of fine-grained evaluation. We introduce ERGeoBench, a diagnostic benchmark for vision-driven embodied geo-localization. ERGeoBench evaluates models under three progressive settings---single-view, panorama-view, and embodied-view---where agents may actively acquire observations through sequential changes in yaw, pitch, and zoom. The benchmark contains 2,207 globally distributed street-view panoramas and measures four complementary capabilities: foundational perception, spatial awareness, common sense reasoning, and geo-localization reasoning. Evaluations of leading proprietary and open-source MLLMs show that current models can infer high-level geographic semantics, but still struggle with fine-grained perceptual operations, metric localization, and spatial consistency across views. We further observe that geo-localization is strongly correlated with the other capability dimensions, suggesting that accurate localization depends on integrated perception, spatial reasoning, and commonsense inference rather than isolated visual recognition. Overall, ERGeoBench provides a unified framework for diagnosing and advancing human-like embodied geo-localization. Project Page: \url{https://kaixuewen.github.io/ERGeoBench/}

Existing deep learning models have demonstrated potential in Diabetic retinopathy (DR) diagnosis, but they still suffer from three key challenges: reliance on single-source inputs, opaque and untraceable reasoning processes, and the absence of a mechanism for result verification. Thus, we propose a provenance-enabled concept-based framework for multi-view DR diagnostic (ProConMV), which integrates DR lesion masks, clinical text and multi-view data, utilizing multimodal prompt analysis and visual-text concept interaction to learn the interpretable multi-source input. During the reasoning stage, the proposed framework introduces lesion concepts for causal reasoning chains combining clinical guidelines, and adds doctor intervention for human-machine collaboration. For dynamic fusion decision and verification in multi-view DR diagnosis, we derive via generalization theory that incorporating each view’s lesion concept uncertainty and grading uncertainty reduces the generalization error upper bound. Accordingly, we design a dual uncertainty-aware module to enable provenance-based verification, ultimately enabling verifiable analysis of DR diagnostic results. Extensive experiments conducted on two public multi-view DR datasets demonstrate the effectiveness of our method. The code will be released at https://github.com/SoY0ung/ProConMV.