Multimodal

Multimodal Large Language Models (MLLMs) can listen and see, but how do audio and visual signals actually travel through the network to shape an answer? Despite their growing role in research and real-world applications, the internal pathways through which audio and visual tokens influence the final prediction remain poorly understood. In this study, we examine audio-visual information flow inside Audio-Visual Large Language Models (AVLLMs), tracing how AVLLMs route, utilize, and integrate audio and visual information across two input configurations, audio-visual video and multiple interleaved audio-visual items. We find that for audio-visual video, AVLLMs follow the sequential information flow pathway established for VLMs and VideoLLMs, with audio and visual contribution flowing along this pathway in proportion to the task's reliance on each modality. In settings with multiple interleaved audio-visual items, this routing shifts to different parallel streams. Furthermore, we demonstrate that audio-visual and other token types can be discarded once their information is transferred to LLM, with minimal impact on the model's prediction or even slight improvement, generalizing across multiple tasks and datasets, enabling more efficient inference. These findings hold across multiple models and scales, Qwen2.5-Omni and Video-SALMONN2 Plus at 3B and 7B scales, leading to hypotheses on why these flow structures emerge. Together, these results deliver the first coherent picture of how AVLLMs orchestrate sound and sight inside the network and lay the groundwork for the next wave of interpretability, design, and efficiency advances in audio-visual and broader MLLMs.

Medical vision-language models (VLMs) are evaluated on public benchmarks whose images and question-answer pairs have been freely downloadable for years, yet reported accuracy assumes these examples were absent from pretraining. We audit open VLMs on SLAKE-En, PathVQA, VQA-RAD, and an auxiliary public OmniMedVQA mirror using four detector families: image-side near-neighbour overlap against PMC-OA-beta, canonical-order exchangeability, cohort-relative Min-K%++ tail enrichment, and cross-model top-K overlap. We find measurable image-side source overlap on SLAKE-En: 19.8% of images are flagged under SigLIP-B-16 and 4.2% under SigLIP-SO400M, while out-of-domain controls produce 0/2000 flags. Manual adjudication shows same-modality, same-projection matches to different patients rather than verified pixel-level duplicates, so we interpret this as source or distributional overlap rather than confirmed per-image memorization. On the text side, Qwen2.5-VL on SLAKE-En shows a canonical-order exchangeability signal that survives ordering ablation and external non-medical baselines. On the OmniMedVQA mirror, exchangeability fires for five medical and general VLMs while BLIP-2 remains clean. In contrast, cohort-relative Min-K%++ tail enrichment and cross-model top-K overlap collapse under an external pre-domain baseline: BLIP-2 reproduces the apparent positive signals despite lacking plausible medical-VQA exposure. We conclude that these cohort-relative detectors are unreliable as standalone membership-inference signals on small medical-VLM cohorts.

Evaluating text-to-music (TTM) systems remains expensive because music impression (MI) and text alignment (TA) scores rely on human mean opinion scores (MOS). Most automatic MOS estimators are trained with point-wise regression or distributional classification. These objectives do not directly optimize rank-based metrics and provide weak geometric constraints for cross-modal coherence. To address these gaps, we propose DeRA-MOS, a decoupled optimization framework for TTM evaluation. For MI, we introduce a batch-aware listwise ranking loss that models relative order within each mini-batch and better aligns with evaluation based on Spearman's rank correlation coefficient (SRCC). For TA, we introduce a score-anchored modality alignment loss that maps human scores to target audio-text similarity and regularizes the latent space before fusion. By effectively mitigating the point-wise training mismatch and modality drift, experiments on MusicEval demonstrate that our decoupled framework yields substantial improvements in both MI and TA ranking metrics, establishing a robust paradigm for large-scale TTM evaluation.

Spatial reasoning is a foundational capability for multimodal large language models (MLLMs) to perceive and operate within the physical world. However, existing benchmarks predominantly rely on passive evaluation (e.g., static VQA) or simulator-specific pipelines, failing to assess general interactive spatial understanding. We introduce SpatialWorld, a unified benchmark designed specifically for evaluating the interactive spatial understanding of multimodal agents in complex real-world tasks. Integrating eight heterogeneous simulation backends under a shared, simulator-agnostic protocol, SpatialWorld features 760 human-annotated tasks across diverse domains (e.g., household routines, travel, social collaboration). Agents must solve tasks under vision-only partial observability, actively gathering egocentric visual evidence and expressing decisions via a unified, text-based action interface native to MLLMs. For reliable evaluation, each task includes a human-validated initial state, a reference trajectory, and a terminal-state verifier. Evaluating 15 advanced agents reveals that robust spatial task solving remains challenging: the strongest model, GPT-5, achieves an average task success rate (TSR) of only 17.4%, while the leading open-source model, Qwen-3.5, reaches 14.1%. Further analysis exposes a clear mismatch between task success and execution efficiency, alongside substantial domain-specific performance variations. These bottlenecks in active exploration and long-horizon planning position SpatialWorld as a rigorous testbed for future spatial agents.

Speech restoration through silent speech interfaces (SSIs) has emerged as a promising assistive technology for individuals with impaired or absent laryngeal voice production. Among non-invasive SSI modalities, surface electromyography (sEMG) and video-based lipreading provide complementary articulatory information, yet their integration for continuous speech synthesis remains underexplored. Moreover, existing multimodal approaches rarely address robustness to modality degradation or temporary sensor failure, limiting their applicability in realistic scenarios. In this work, we propose a masked multimodal speech synthesis framework that jointly leverages sEMG and lipreading signals through modality masking during training. Under multispeaker settings, the proposed approach reduces word error rate by up to 14 absolute percentage points compared to the strongest unimodal baseline. Experimental results not only show that masking strategies are critical for these performance gains and robustness under low-bitrate conditions, but also that they generalize better than degradation-specific data augmentations in the presence of modality absence conditions. Phone-level analyses further reveal complementary contributions across modalities, with particularly strong benefits for vowels and for specific consonant groups. Overall, these findings demonstrate the effectiveness and robustness of masked multimodal integration for silent speech synthesis, although adaptation to laryngectomized speakers remains an open research challenge.

Multimodal large language models (MLLMs) achieve strong results on visual reasoning benchmarks, but answer accuracy alone does not indicate whether a model relied on the correct visual evidence. This gap is particularly important in multi-view driving scenes used for autonomous driving, where a model can produce a plausible answer while grounding it in the wrong camera view. We introduce a multi-view visual question answering benchmark for evaluating evidence-source identification: given six synchronized NuScenes views and a question, the model must identify the supporting camera view and answer the question. The benchmark contains 122 conflict-centric question-answer pairs from 73 scenes, spanning causality, counterfactual reasoning, and intent prediction. View labels are proposed by an automatic conflict-mining pipeline and manually verified by annotators. We evaluate three settings: camera-view selection, oracle QA given the golden view, and joint prediction in which the model selects a view and answers in one pass. Answers are evaluated in both multiple-choice and free-form formats, using exact match for structured predictions and an LLM judge for free-form responses. By explicitly separating visual-source identification from answer correctness, the benchmark exposes grounding failures that answer-only evaluation misses.

Chain-of-Thought (CoT) improves the performance of Large Language Models (LLMs) and has been extended to Multimodal Large Language Models (MLLMs). More recent work further moves from text-based multimodal reasoning toward interleaved-modal reasoning, where intermediate steps can incorporate both textual rationales and visual evidence. In this work, we propose a bolder and more ambitious idea: could images alone serve as the reasoning medium for both language and multimodal tasks? To explore this, we propose optical reasoning, which treats images as a standalone reasoning medium. We instantiate this concept with two variants: typographic-based optical reasoning, which optimizes visual layouts for compact rationale rendering, and graphical-based optical reasoning, which composes text and graphical elements into structured visual rationales. Across mathematical, scientific, and interleaved-modal reasoning benchmarks, optical reasoning can match or even exceed traditional text reasoning while reducing reasoning tokens by an average of 28.57% on language tasks and 16% on multimodal tasks, achieving 1.96 times the token efficiency of text reasoning. These results show that images can effectively and efficiently encode rationales while providing a unified visual canvas for reasoning.

Open-domain open-vocabulary detection (ODOVD) requires detectors to generalize to both novel categories and unseen domains, making it more challenging than open-vocabulary detection. Existing methods typically train open-vocabulary detectors together with domain generalization modules from scratch, leading to high training cost. we propose ExDet, a lightweight category-domain collaborative generalization framework for ODOVD that enhances the cross-category and cross-domain generalization of existing detectors. ExDet consists of Text-Guided Extrapolation (TGE), a lightweight Detector-Compatible Rectification (DCR) module, and ExRPN. Specifically, TGE exploits the DeltaSpace property of vision-language models (VLMs) to infer category- and domain-aware proxy visual prototypes from text. DCR is learned from the TGE-generated prototypes in a detector training-free and real-data-free manner, and is inserted after the classification head at inference to rectify representations toward a detector-compatible source-domain visual distribution, thereby enhancing classification for targets from novel categories and unseen domains. ExRPN recalibrates proposal scores by combining semantic similarity with RPN confidence, improving recall for novel and domain-shifted objects while providing better support for subsequent classification and DCR. ExDet achieves SOTA performance on OD-LVIS, OV-LVIS, Objects365, and MSOSB.

Synthetic aperture radar (SAR)-assisted optical cloud removal aims to recover surface information obscured by clouds in optical remote sensing images by exploiting complementary SAR observations. Existing multimodal fusion methods typically rely on direct spatial concatenation and pixel-wise supervision, which can propagate SAR speckle noise into optical reconstruction and lead to over-smoothed results. To address these limitations, we propose an Information Bottleneck-driven High-Fidelity Network (IB-HFN) for SAR-assisted optical cloud removal. IB-HFN employs a dual-stream backbone to preserve modality-specific representations before deep semantic fusion, thereby mitigating premature cross-modal contamination. At the fusion stage, we introduce a Spatial Information Bottleneck Fusion module that compresses SAR features through a channel-wise variational information bottleneck to suppress unstructured speckle noise. In parallel, a local-global gating mechanism predicts clear-sky regions and routes reliable optical details through a Dirac-initialized skip connection, decoupling noise suppression from texture preservation. We further develop a joint optimization strategy that integrates feature-level bottleneck regularization with image-level constraints on reconstruction accuracy, structural consistency, spectral fidelity, and contrastive sharpness. A dynamic weighting schedule balances these objectives to stabilize training and reduce hazy artifacts. Experiments on the SEN12MS-CR dataset under challenging spatio-temporal splits demonstrate that IB-HFN achieves superior structural preservation and spectral fidelity over existing methods.

Omni-modal retrieval promises a single embedding space for text, image, video, document, and audio inputs, but building such a unified retriever is difficult since these modalities differ in data distribution, architecture, and optimization dynamics. In this work, we present Conan-embedding-v3, a decouple--fuse--recover framework for omni-modal retrieval. Conan-embedding-v3 first trains modality specialists independently and fuses their task vectors into a single dense backbone, a strategy we call Decoupled Specialist Fusion. We show that this fusion composes visual, video, and document retrieval capabilities, but also exposes a failure mode for projector-based modalities: when audio is attached through an external encoder and projector, fusing the backbone leaves the projector calibrated to the audio-specialist backbone, causing a large audio retrieval regression despite copying all audio-specific modules unchanged. We call this failure Projector Drift. To repair it, Conan-embedding-v3 applies Projector Recovery (i.e., full-parameter fine-tuning of the projector while keeping the backbone frozen) followed by balanced multi-modal rehearsal. The resulting model supports these retrieval pathways in one backbone, achieving 74.9 scores on MMEB while obtaining 55.61 on the 30-task MAEB audio suite.

The rapid development of pretrained foundation models has enabled more general image segmentation. Multimodal large language models (MLLMs) have been widely explored for image segmentation with complex queries that require high-level reasoning. Despite promising progress, existing methods are often constrained by limited training data and the gap between MLLMs and mask generation modules. To better transfer MLLMs' perception and reasoning ability to complex reasoning-based segmentation tasks, we propose a two-stage framework Rea2Seg for mask generation and selection. Specifically, the framework first identifies potential regions as candidate masks based on the attention maps of a segmentation MLLM. It then employs an MLLM to reason over the question and candidate masks and assign scores to each mask. The final segmentation result is obtained by reranking the candidates and selecting the highest-scoring mask, reformulating image segmentation as candidate discovery followed by discriminative mask selection. We also notice that a large portion of questions in existing benchmarks focus on commonsense reasoning, and these questions usually do not fully require joint visual observation and reasoning. To address this issue, we introduce a new benchmark called ReasonSeg-SGDR that comprehensively evaluates a model's perception, grounding, and reasoning abilities across multiple dimensions, including discriminative recognition, spatial reasoning, geometric reasoning, and multi-step reasoning, with fine-grained mask generation. In addition, we collect training data to enhance MLLMs' ability to jointly understand multimodal queries and candidate masks, and to assign scores through reasoning. Experimental results on the proposed benchmark and ReasonSeg demonstrate the effectiveness of the unified mask generation and selection framework.

Parkinson's disease (PD) is a progressive neurodegenerative disorder that frequently causes speech impairments associated with hypokinetic dysarthria. As speech production relies on the precise coordination of complex neuromuscular mechanisms, speech analysis has emerged as a promising non-invasive and cost-effective biomarker for early PD detection. Recent deep learning approaches have shown encouraging results; however, most existing methods rely on a single speech representation, potentially overlooking complementary pathological information encoded across different feature spaces. In this work, we propose a multi-branch deep learning framework for automatic PD detection from speech. Each recording is segmented into 5-second chunks and represented using three complementary modalities: Log-Mel spectrograms, MFCCs, and HuBERT embeddings extracted from raw waveforms. The spectrograms are processed using a pre-trained ResNet-18 encoder, MFCC sequences are modeled through a BiLSTM network, and raw speech is encoded using a pre-trained HuBERT model. To effectively integrate these heterogeneous representations, we introduce a context-guided cross-modal attention mechanism that dynamically weights temporal HuBERT embeddings according to the global acoustic context derived from the spectrogram and MFCC branches. Experiments conducted on the publicly available Spanish PC-GITA corpus under strict speaker-independent 5-fold cross-validation demonstrate the effectiveness of the proposed approach. The proposed architecture achieves an accuracy of 91.51%, an F1-score of 91.24%, and an AUC of 95.97%. Furthermore, ablation studies confirm the contribution of both the proposed context-guided cross-modal attention mechanism and the integration of complementary speech representations. These findings highlight the potential of heterogeneous speech modeling for robust and clinically reliable PD detection.

Two-view correspondence learning aims to distinguish true correspondences (inliers) from false ones (outliers) in image pairs by leveraging their underlying differences. Existing methods mainly rely on coordinate-based geometric consistency. However, they often struggle with pseudo-consistent outliers in scenes containing repetitive structures, textureless regions, or locally similar geometric patterns. To address this limitation, we propose TriMatch, a multi-source feature fusion framework for two-view correspondence learning, which consists of two parts: feature extraction and feature refinement. In feature extraction, TriMatch jointly extracts geometric, texture semantic, and structural semantic features to provide complementary evidence for correspondence discrimination. To bridge the gap between semantic and geometric features, texture and structural semantic features are aligned with geometric features through dedicated Texture-Geometric Alignment and Structural-Geometric Alignment modules, respectively. We further introduce a Semantic-Guided Correspondence Modulation module, which modulates geometric features using semantic information to suppress geometrically plausible but semantically inconsistent correspondences. In feature refinement, a Hierarchical Semantic-Enhanced Correspondence Refinement strategy progressively models correspondence dependencies and recalibrates multi-context feature responses, enabling more reliable inlier-outlier discrimination. Extensive experiments demonstrate the effectiveness, robustness, and generalization capability of TriMatch.

Understanding and reasoning over abstract visual content remains a challenge for current multi-modal large language models (MLLMs). In this paper, we explore a novel abstract data type termed complex visual query (CVQ), designed to probe symbolic and abstractive reasoning, which is a critical yet underexplored dimension of human-like neuro-symbolic reasoning for MLLMs. We present a comprehensive investigation from three perspectives: \textbf{Data $\times$ Paradigm $\times$ Exploration}. Specifically, we propose a scalable pipeline for synthesizing CVQs grounded in large-scale multi-modal knowledge graphs, generating a diverse dataset encompassing 14 distinct query types via systematic combinations of first-order logic operators. We further introduce a two-stage training framework that progressively equips MLLMs with robust visual reasoning capabilities. We conduct extensive experiments to rigorously evaluate MLLMs across multiple dimensions, including reasoning performance on CVQs, as well as cross-task and cross-scenario generalization. We believe our work opens new perspectives and avenues for advancing the reasoning frontiers of MLLMs.

Hyperspectral object tracking (HOT) leverages the rich spectral information provided by hyperspectral videos (HSVs), offering substantial potential for object tracking. However, efficiently extracting and exploiting spectral information from redundant spectral bands remains a fundamental challenge, which severely limits model generalization and tracking performance. Moreover, in dynamic scenes, targets often experience drastic appearance variations due to factors such as occlusion and illumination changes. These variations lead to large deformations between the current frame and the template. Such discrepancies pose major challenges for existing temporal modeling approaches. In this work, we propose VLHTrack, a novel hyperspectral vision-language (VL) joint tracking framework. Specifically, we incorporate language priors to address the fundamental challenge of spectral redundancy by designing a Language-Guided Band Selection Module (LBSM). By leveraging Large Language Model (LLM) descriptions, LBSM establishes a semantic-to-spectral mapping that mitigates redundancy and accentuates discriminative spectral features. A Multi-Modal Vision-Language Fusion Module is then employed to seamlessly integrate visual and linguistic embeddings, harnessing their complementary advantages to learn coherent cross-modal representations. To address target deformation in long-term sequences, we propose a dynamic update template feature strategy implemented via the Dynamic Template Update with Mamba (DTUM) module. By leveraging selective state space modeling, DTUM learns inter-frame dependencies to update template feature, ensuring efficient template feature evolution guided by temporal context. Experiments on HOT2023 and HOT2024 demonstrate that VLHTrack outperforms state-of-the-art (SOTA) methods.

Multimodal affective analysis aims to understand human sentiment and emotion by jointly modeling heterogeneous modalities such as text and images. However, multimodal models often fail to consistently outperform strong text-only baselines, with performance varying significantly across fusion strategies. In this work, we identify representation misalignment between independently pretrained modality encoders as a key bottleneck for effective multimodal learning, and show through controlled experiments that alignment prior to fusion is often more important than fusion complexity. To address this issue, we propose a unified multimodal affective analysis framework that leverages vision-language models (VLMs) to convert visual content into structured textual descriptions, projecting heterogeneous modalities into a shared linguistic space and enabling interpretable text-centric reasoning. To further improve robustness, we introduce a hybrid learning strategy that combines semantic token selection with a batch-level uniformity regularization objective, encouraging a more dispersed and stable global feature space while mitigating noise introduced by VLM-generated descriptions. Experiments on multiple multimodal sentiment and emotion benchmarks show that our method consistently outperforms strong unimodal and multimodal baselines, achieving state-of-the-art performance. Our analysis further highlights the critical role of representation alignment in multimodal affective learning.

Multimodal large language models (MLLMs) commonly inherit the deep, symmetric Transformer backbone designed for unimodal text modeling, and apply the same computation uniformly to image and language tokens. This design overlooks a key modality asymmetry: image and text tokens differ substantially in information density, redundancy, and required reasoning depth. Through a layer-wise analysis of LLaVA-1.5, we observe that vision tokens tend to saturate in the middle layers. Specifically, text-to-image attention decreases from 0.68 at layer 0 to 0.07 by layer 4, and stabilizes near 0.04 after layer 18, whereas text tokens continue to benefit from deep semantic processing. These findings suggest a mismatch between architectural symmetry and depth-asynchronous modality evolution, resulting in redundant visual computation and possible drift in perceptual representations during deep task-specific adaptation. Motivated by this, we propose Dual-Path Vision Token Routing (DPVR), a modality-asymmetric routing framework for efficient MLLMs. Its core instantiation, DPVR-LF (Late-Layer Fusion), routes vision tokens at the saturation point into a one-layer trainable side branch, runs a thirteen-layer text-only forward that skips image positions in the deep stack, and re-fuses the visual and textual streams only at the final layer. With approximately 3% trainable parameters, DPVR-LF preserves competitive multimodal performance on standard benchmarks while reducing visual computation in the deep Transformer stack. The results challenge the conventional assumption that vision tokens must traverse all deep language-model layers, and indicate that a single late fusion layer can be sufficient for maintaining strong perceptual competence in LLaVA-style MLLMs.

Privacy risks in text-only Large Language Models (LLMs) are well studied, particularly their tendency to memorize and leak sensitive information. However, Multi-modal Large Language Models (MLLMs), which process both text and images, introduce unique privacy challenges that remain underexplored. Compared to text-only models, MLLMs can extract and expose sensitive information embedded in images, posing new privacy risks. We reveal that some MLLMs are susceptible to privacy breaches, leaking sensitive data embedded in images or stored in memory. Specifically, in this paper, we (1) introduce MM-Privacy, a comprehensive dataset designed to assess privacy risks across various multi-modal tasks and scenarios, where we define Disclosure Risks and Retention Risks. (2) systematically evaluate different MLLMs using MM-Privacy and demonstrate how models leak sensitive data across various tasks, and (3) provide additional insights into the role of task inconsistency in privacy risks, emphasizing the urgent need for mitigation strategies. Our findings highlight privacy concerns in MLLMs, underscoring the necessity of safeguards to prevent data exposure. Our dataset and code can be found here.

Comprehensive estimation of dietary micronutrients from food images could improve clinical nutrition care, but training such models requires large multimodal datasets linking diverse foods to complete nutrient profiles. We first show that existing multimodal large language models (MLLMs), including leading proprietary models, are unreliable for this task. Across five model families and four independent evaluation benchmarks (ASA24, SNAPMe, FNDDS, and NutriBench), models frequently abstained or returned statistically implausible values. To address this gap without costly expert annotation, we repurposed a decade of population-scale 24-hour dietary recalls as structured prompts for text-to-image generation. This pipeline produced a synthetic corpus of about 1.1 million image-description-nutrient triplets, each pairing a generated food image with a complete 65-nutrient label. To our knowledge, this is the largest synthetic food-image corpus with comprehensive micronutrient annotation planned for public release upon publication. Fine-tuning Qwen3-VL (2B/4B/8B/30B) and GLM-4.6V-Flash on this corpus yielded NutriMLLM, the first family of vision-language models specialized for comprehensive dietary micronutrient estimation. We evaluate these models with a four-component framework that separately measures abstention, hallucination, overall usability, and per-nutrient numerical accuracy. On real food images, every NutriMLLM variant achieved near-complete coverage across all 65 nutrients, and the largest variant matched or exceeded proprietary baselines (GPT-5, Gemini 3, and Claude Sonnet 4.5) in accuracy on most nutrients. These results show that recall-driven synthetic supervision can make image-based comprehensive micronutrient estimation a tractable engineering problem and support dietary assessment, personalized nutrition guidance, and population-scale micronutrient surveillance.

In many binary segmentation tasks, most multimodal methods rely on fixed feature concatenation for cross-modal interaction and straightforward decoder designs dominated by low-frequency semantics. However, they ignore two key challenges: one is the lack of an adaptive mechanism to handle modality discrepancies and complementarity, and the other is the absence of an efficient decoding strategy to balance both high- and low-frequency representations. In this work, we propose a simple yet general multimodal binary segmentation framework, termed DifferSeg, to address both problems simultaneously. With the help of the differential perception fusion (DPF) module, DifferSeg employs learnable differential operators to adaptively align multimodal features and enhance their complementarity through residual fusion, effectively mitigating modality mismatch and fusion redundancy. In addition, we design a frequency-guided decoder (FGD) that builds cross-frequency interactions and multi-path upsampling to maintain consistency between detailed high-frequency structures and semantic low-frequency representations, ensuring fine-grained boundary recovery and noise suppression. % Benefiting from these designs, DifferSeg can be easily generalized to diverse binary segmentation tasks, including both natural and medical modalities. Without bells and whistles, it consistently surpasses 67 state-of-the-art methods across 29 public datasets involving 18 downstream tasks, demonstrating superior generalization and segmentation accuracy.Code and pretrained models will be available at the Link.

Reinforcement Learning with Verifiable Rewards (RLVR) has become an effective paradigm for improving the reasoning capability of Large Vision-Language Models (LVLMs). However, existing RLVR methods primarily rely on trajectory-level outcome rewards, which assign identical learning signals across all generated tokens. This coarse-grained credit assignment is fundamentally mismatched to multimodal reasoning, where only a sparse subset of tokens is causally grounded in visual evidence. Consequently, these pivotal perceptual tokens receive weak supervision and are often overwhelmed by language priors or reasoning-template tokens. To address this limitation, we propose Perception-Reinforced Policy Optimization (PRPO), a token-level reinforcement learning framework that explicitly identifies and reinforces pivotal perceptual tokens within long-horizon multimodal reasoning trajectories. PRPO introduces Robust Visual Dependency (RVD), a principled metric that identifies tokens whose predictions are both visually grounded and perturbation-stable, filtering out brittle or noisy visual tokens. Based on RVD, we further propose Perceptual Advantage Reshaping (PAR), a token-level credit assignment technique that amplifies perceptually informative tokens while preserving stable gradients for non-perceptual tokens. Extensive experiments on seven multimodal reasoning benchmarks demonstrate that PRPO consistently outperforms strong LVLM baselines across both 3B and 7B model scales, achieving average gains of 23.3% and 21.1%, respectively. PRPO achieves state-of-the-art performance with improved training efficiency and stronger cross-task generalization. Our findings highlight the importance of fine-grained credit assignment for scalable multimodal reinforcement learning.

Multimodal Large Language Models (MLLMs) face a significant inference bottleneck due to the quadratic computational cost of self-attention over long visual token sequences. However, we identify a critical inefficiency in current architectures: Visual Attention Saturation. Our analysis reveals that visual tokens rapidly establish their spatial structure and intra-modal relationships in early layers, rendering visual-to-visual self-attention in deeper layers computationally redundant. Conversely, Feed-Forward Networks (FFNs) in these layers remain essential for projecting visual features into the evolving textual semantic space. Leveraging this insight, we present Visual-Skip (V-Skip), a training-free inference paradigm that decouples spatial interaction from semantic evolution. Rather than discarding tokens, V-Skip imposes block-wise structured sparsity by selectively bypassing saturated visual self-attention modules. Furthermore, recognizing that varying downstream tasks demand distinct reasoning depths, V-Skip employs a lightweight, few-shot calibration to dynamically route the task-optimal sparsity path. Extensive experiments demonstrate that V-Skip effectively bypasses redundant vision attention to achieve block-wise sparsity, maintaining a 94.16% to 100.31% performance retention across diverse MLLMs. Ultimately, we prove that to reason more effectively, models do not need to discard what they see -- they simply need to "look less" at the right depth.

Multimodal language models are typically evaluated through external behavior: selecting the correct image--text match, rejecting unsupported captions, or answering visual queries correctly. However, correct behavior alone does not show that the model's internal decision state remains stable under controlled semantic stress. We study this gap through S$^3$E (Structured Semantic Stress Evaluation), a framework for analyzing behavior-internal decoupling in multimodal language models. S$^3$E uses a positive-anchored A/B forced-choice setup in which an image-supported caption is contrasted against semantic stress candidates under both original and swapped option orders, while hidden states are extracted at the pre-answer decision state. We focus on strict-correct trials, where the model consistently selects the correct caption across both orders. Rather than treating arbitrary hidden-state variation as evidence of instability, we measure whether semantic-conflict candidates induce excess decision-state displacement relative to meaning-preserving controls. Across Qwen3VL, Gemma3, and InternVL3, semantic stress consistently produces positive selected-layer excess displacement over lexical controls despite correct forced-choice behavior, while comparisons against random negatives are model-dependent. We interpret this as a scoped decision-state stress-sensitivity signal rather than evidence of downstream failure or hallucination. Our results suggest that forced-choice correctness alone is not a sufficient certificate of invariant internal decision geometry.

While multimodal integration significantly improves computer vision models, deploying them incurs prohibitive inference costs and requires scarce, perfectly paired datasets. Recent methods address this data bottleneck by synthesizing missing modalities via generative AI, yet they introduce a severe inefficiency: the Decode-Encode Loop. Specifically, information-rich generative latents are decoded into noisy raw signals, forcing the downstream classifier to waste capacity re-encoding them. To bypass this bottleneck, we propose Direct Latent Augmentation (DLA), utilizing undecoded generative latents directly as privileged information. Furthermore, to transfer this dense knowledge to a purely visual student, we introduce Multilayer Explicit Simulated Synesthesia (MESSy). Instead of enforcing rigid representation matching, which forces the student to distort its native visual features to accommodate complex multimodal topologies, MESSy uses a predictive objective to safely internalize these physical priors. Empirical results demonstrate that our framework significantly outperforms raw data augmentation and traditional distillation. Ultimately, our approach yields highly accurate unimodal students with "synesthetic" latent structures that are inherently aligned with modalities they have never directly observed.

Multimodal large language models (MLLMs) have made substantial advancements in video understanding, yet the reliability of their responses remains underexplored. This work presents a diagnostic study of absent answer detection for MLLMs in video understanding, where the correct answer is deliberately excluded from the candidate set and a reliable model is expected to recognize that no valid option exists. We evaluate the absent answer detection behavior under three settings: multiple-choice questions augmented with an ``None of the Above'' option, open-ended generation with a detection instruction, and standard evaluation without any guidance. Across a diverse set of models and benchmarks, we find that MLLMs overwhelmingly select plausible distractors rather than detecting the absent answer. This failure is more pronounced in temporal reasoning tasks and worsens with denser frame sampling. We further explore chain-of-thought prompting as a mitigation strategy and find that while it substantially improves detection rates, performance remains unsatisfactory, suggesting that prompting-based strategies alone are insufficient to fully address this limitation. These findings expose a systematic failure in absent answer detection and highlight the need for explicit detection mechanisms in multimodal systems.

Test-time Scaling (TTS) has emerged as a pivotal research direction for enhancing model performance by dynamically allocating computational resources during inference. Recent advancements have adapted this paradigm to Multimodal Foundation Models (MFMs), unlocking their potential in multimodal reasoning and generation. Despite rapid progress, the field lacks a systematic survey and unified theoretical framework to delineate the developmental landscape of multimodal TTS. To bridge this gap, we present the first comprehensive review of TTS research for MFMs, proposing a unified taxonomic framework that categorizes existing methodologies into three distinct strategies: sampling-based, feedback-based, and search-based approaches. We further summarize representative applications and benchmarks commonly utilized to evaluate multimodal TTS capabilities in generation and reasoning tasks. Finally, this survey discusses open challenges and outlines future research directions, providing a systematic roadmap for subsequent studies in this rapidly evolving field.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable success in visual understanding, yet their performance degrades significantly under real-world visual corruptions. While existing robustness enhancement approaches exist, they are limited: black-box feature alignment lacks interpretability, and white-box text-based reasoning cannot restore lost pixel-level details. This work investigates a fundamental research question: Can MLLMs recover corrupted visual content by themselves? To address this, we propose Robust-U1, a novel framework that equips MLLMs with explicit visual self-recovery capability for robust understanding. The approach comprises three core stages: supervised fine-tuning for initial reconstruction, reinforcement learning with dual rewards (pixel-level SSIM and semantic-level CLIP similarity) for aligning high visual quality, and multimodal reasoning that jointly considers both the corrupted input and the recovered image. Extensive experiments demonstrate that Robust-U1 achieves state-of-the-art robustness on the real-world corruption benchmark and maintains superior performance under adversarial corruptions on general VQA benchmarks. Analysis confirms that high-quality visual recovery directly enhances reasoning performance, establishing self-recovery as a critical mechanism for robust visual understanding. The source code is available at https://github.com/jqtangust/Robust-U1.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a leading paradigm for enhancing visual reasoning in Multimodal Large Language Models (MLLMs). However, existing RLVR methods optimize primarily for the reasoning outcome, fundamentally overlooking the fine-grained cross-modal coordination required during the generation process. Through token-level analyses and controlled interventions, we reveal that during Chain-of-Thought (CoT) reasoning, MLLMs frequently fail to dynamically alternate between extracting visual evidence and synthesizing textual context-a coordination breakdown that is causally linked to reasoning failures. Motivated by these findings, we propose DyCo-RL, which integrates dynamic cross-modal coordination into RLVR optimization. Specifically, DyCo-RL uses the Fisher-Rao geodesic distance to measure within-modality attention shifts, assigning tokens to either visually-oriented or text-oriented functional roles. It then evaluates the alignment between a token's actual attention allocation and its assigned role, leveraging this score for alignment-guided advantage reweighting during policy optimization. Extensive experiments demonstrate that the algorithm-agnostic DyCo-RL, when applied to Qwen2.5-VL-3B/7B, consistently improves four representative RLVR algorithms across seven benchmarks spanning visual-centric and mathematical reasoning.

Multimodal clinical learning is increasingly important for integrating diverse patient data, including imaging, text, and personalised health records. However, it faces two fundamental challenges: i) modality missingness, where arbitrary subsets of modalities are unavailable at a given patient visit, ii) longitudinal dynamics, where the diagnostic significance of an observation depends on the patient's evolving disease trajectory over time. Existing methods address these challenges in isolation: missing-modality frameworks treat each visit as an independent static snapshot and discard temporal context, while longitudinal models often assume complete modality availability and degrade under systematic modality incompleteness. We propose LongMoE (Longitudinal Mixture-of-Experts), the unified framework to jointly address both challenges. LongMoE combines a context-aware imputation module with an attentional tokenization module that captures frequency-domain temporal patterns across irregular visit sequences, a trajectory-aware encoder for modeling disease progression, and context-conditioned Sparse MoE routing for patient-specific expert selection. Experiments on ADNI, OASIS-3, and MIMIC-IV show that LongMoE improves robustness under missing or weak contemporaneous modalities and remains competitive in full-modality settings, establishing a strong foundation for longitudinally-aware multimodal clinical learning.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in open-world reasoning and understanding. However, a critical ambiguity persists: it remains unclear whether these models genuinely synthesize cross-modal information to construct physically grounded reasoning chains, or if they merely exploit strong language priors to mask single-modality reliance, thereby hallucinating advanced multimodal capabilities. Motivated by this, and to rigorously mitigate language modality bias and shortcuts, we propose a novel multimodal Chrono}logical Physical Dynamics Reasoning Benchmark ChronoPhyBench, which unifies next state prediction with Visual Question Answering (VQA) paradigms by conditioning on historical video context and textual captions to enforce models to deduce subsequent physical states through both single image selection and the inherently more complex task of multiple frame chronological sorting. Concurrently, we construct a large-scale multimodal reasoning dataset curated using the ChronoPhyBench criteria, comprising over 10,000 long-form videos paired with meticulously annotated captions, totaling 5M tokens. Our experimental evaluations reveal a stark contrast to conclusions drawn by previous benchmarks. The capacity of current open-source models to perform physically grounded multimodal reasoning remains in its infancy. Ultimately, this work seeks to systematically stress-test the reasoning capabilities of multimodal models, quantify hallucination rates, and advance the development of Physical AI, thereby providing the community with a robust and transparent evaluation framework toward Artificial General Intelligence (AGI).

Large-scale Visual-Language Models (LVLMs) have achieved remarkable success in natural visual tasks, yet their application to industrial defect detection remains challenging due to two fundamental limitations: (i) the scarcity of large-scale industrial datasets that cover diverse defect categories across multiple domains, and (ii) the reliance on manual prompts (points, boxes, masks) that introduce subjective noise and lack text-visual interaction for fine-grained understanding. To address these challenges, we introduce a Large-Scale Multi-Modal Industrial Open-Closed benchmark (MMIOC-1M) containing over one million samples across $14$ super-categories, $29$ industrial scenes, and $351$ defect subcategories. To our knowledge, MMIOC-1M is the first unified largest benchmark supporting both open-vocabulary and closed-set industrial detection, providing valuable pre-training data for LVLMs in industrial scenarios. Furthermore, we propose a Refined Text-Visual Prompt Network (RTVPNet) that incorporates three key innovations: (1) an expert-assisted domain projection mechanism that enables rapid adaptation of general vision models to industrial domains, (2) an energy-based sparse sampling strategy that automatically generates refined visual prompts without manual intervention, and (3) a bidirectional text-visual interaction module that enhances cross-modal semantic alignment and understanding. Extensive experiments demonstrate that RTVPNet achieves state-of-the-art performance on MMIOC-1M, LVIS, and COCO benchmarks while maintaining computational efficiency. The dataset and code are available at https://github.com/hellozzk/MMIO.

This paper presents our system description for the 2nd Workshop on Multimodal Augmented Generation via MultimodAl Retrieval (MAGMaR). Addressing the critical challenges of cross-lingual long-video comprehension, strict persona adherence, and zero-hallucination temporal grounding, we propose a fully training-free, two-stage cascaded Video RAG pipeline. Our architecture strategically decouples semantic retrieval from cognitive logical reasoning through a modality-aware division of labor. In the first stage, a high-recall semantic pre-fetching module employs dense retrieval using only high-fidelity visual summaries and global text descriptions, explicitly isolating noisy modalities (e.g., OCR and ASR) to maintain a pristine vector space. In the second stage, an Adaptive, Iterative, and Reasoning-based (A.I.R.) filtering agent, powered by a commercial Large Language Model (LLM), performs fine-grained cognitive reranking. The agent re-incorporates full multimodal contexts to enforce strict logical alignment with user personas, effectively pruning semantically similar but logically irrelevant candidates. Finally, a Prompt Sculpting mechanism constrains the generator to synthesize the distilled subset into strictly formatted JSON responses with exact chunk-level citations. Evaluated on the RAG track, our resource-aware approach shows exceptional precision in both information retrieval and persona-conditioned generation.

When a multimodal large language model answers a visual reasoning question correctly, is the prediction actually supported by the task-critical visual evidence? Correct answers can coexist with flawed reasoning, making accuracy alone an incomplete test of grounding. We introduce VisualFLIP, a paired benchmark with 1,374 images arranged as same-question perturbation pairs across cardinality, attribute, spatial, and logic tasks. Each pair keeps the question fixed but minimally changes the evidence so the gold answer deterministically flips. We evaluate 24 MLLMs with pair accuracy, which requires solving both sides of a pair, and Collapse Rate (CR), which measures how often a model that solves at least one side repeats the same non-empty answer for both images. Together, these metrics show that paired correctness and evidence dependence are related but distinct: capable models can still fail to update after task-critical visual changes, and collapse becomes more severe for some models when the edited image follows an earlier answer in a sequential setting. Further details are available on our project page: https://didizhu-judy.github.io/VisualFLIP/

Multimodal reasoning requires a path that retains integrity over a wide range of constraints, from visual grounding to logic consistency. However, the current Process Reward Models focus on heuristically defined rewards that equally weigh these factors, which may lead to the concealment of individual dimension failures by the dominating factors, without guaranteeing the validity of the reasoning process in general.

Video understanding is being rapidly transformed by multimodal large language models (MLLMs), as research moves from short clips to long, multimodal, and knowledge-intensive video scenarios. These scenarios require models to handle sparse evidence, long-range dependencies, multimodal alignment, and reliable inference under limited computational budgets. This work presents a human-view perspective on LLM-based video understanding, organized around three functional abilities: watching, remembering, and reasoning. Rather than treating video tasks as isolated benchmarks, this view provides a unified structure for analyzing how video MLLMs acquire evidence, preserve context, and produce grounded outputs. We introduce a formulation that characterizes video understanding systems by their perceptual representations, memory states, reasoning traces, and final predictions. Based on this formulation, we identify challenges in spatio-temporal perception, efficient long-video processing, memory modeling, streaming understanding, and faithful reasoning. Representative methods are organized by their roles in video MLLM systems. Watching covers fine-grained, comprehensive, audio-visual, and efficient perception. Remembering includes offline and streaming memory, while reasoning covers text-only reasoning and thinking with videos. We further examine application domains such as egocentric, sports, instructional, medical, and narrative videos, and cover training datasets and evaluation benchmarks across task types, supervision formats, modalities, and capability dimensions. Finally, we outline open problems and future directions for scalable, memory-aware, and evidence-grounded video intelligence. Related works will be continuously traced at https://github.com/marinero4972/Awesome-HumanView-VideoUnderstanding.

Language agents are increasingly deployed over accumulating multimodal information, yet existing benchmarks assume a human-human form with sparse visuals and straightforward content, evaluating neither reasoning over authentic multimodal file interaction nor the interpretation of concealed user information. We therefore introduce M$^3$Exam, a query-centric multimodal conversational memory benchmark built on realistic user-agent interaction, with multi-dimensional evaluation spanning cross-modal grounding and implicit information inference. Benchmarking MLLMs and memory systems reveals persistent gaps in cross-modal grounding, cross session reasoning, and the efficiency cost of accumulating multimodal context. We further propose M$^3$Proctor, a multimodal memory method that detects query modality bias and consumes raw visual sources only on demand, improving accuracy by 13% while cutting index-construction time and retrieved tokens by over 70%.

Multimodal large language models (MLLMs) have raised new privacy challenges. On the data side, user-provided inputs often include unpredictable sensitive information; while on the downstream task side, model reasoning depends on rich visual context that may itself be privacy-sensitive. Existing privacy protection methods, however, rely on predefined sensitive categories and fixed obfuscation strategies, struggling to tackle such challenges in MLLMs. To address this dilemma, we propose Anchored Privacy Drifting (APD), a training-free method that drifts privacy-sensitive elements toward semantically equivalent alternatives while anchoring contextual cues to the source image. To systematically evaluate this dual objective of privacy protection and contextual preservation, we introduce AdaptShield, a comprehensive benchmark covering 22 privacy categories, which combines conventional privacy metrics with MLLM-based assessments of contextual utility. Extensive experiments show that our method achieves balanced improvements in both privacy sanitization and content retention, with average gains of 10.4% on textual categories and 8.5% under MLLM-based evaluation across four MLLM series, i.e., Qwen2.5, Qwen3, InternVL3, and InternVL3.5.

World Action Models (WAMs) offer a promising approach to embodied intelligence, yet existing methods rely heavily on video prediction as action priors and lack adaptive multimodal reasoning, limiting their effectiveness on long-horizon, complex tasks. We observe that WAMs require different multimodal reasoning modes under different execution contexts: textual reasoning is essential during task transitions to guide high-level action prediction, while visual reasoning is critical during fine-grained manipulation for precise control. Motivated by this observation, we propose \textbf{AdaWAM}, a world action model with adaptive multimodal reasoning abilities. AdaWAM integrates a lightweight dynamic router that autonomously triggers textual or visual reasoning as needed during task execution. Experiments on both simulated and real-world embodied tasks show that AdaWAM substantially improves inference efficiency while outperforming state-of-the-art embodied policies. Codes and demos are available at: https://adawam.github.io/.

Multi-modal approaches used for 3D CAD generation require substantial computational resources, necessitating efficient training. To address this, we propose GuideCAD, which leverages semantically rich visual-textual representations having only a small number of trainable parameters to generate 3D CAD models. Specifically, GuideCAD uses a mapping network that converts image embeddings into prefix embeddings, enabling a pretrained large language model to integrate visual and textual information. As a result, a transformer-based decoder predicts the construction sequence using the visual-textual embeddings in order to generate the 3D CAD model. For experimental evaluation, we construct a new dataset, referred to as GuideCAD, which consists of text-image pairs. Each pair includes a text prompt that represents a 3D CAD construction sequence and its corresponding 3D CAD image. Our experimental results show that GuideCAD generates comparably high-quality 3D CAD models while using approximately four times fewer parameters and achieving twice the training efficiency compared to fine-tuning approaches. We have released the source code and dataset for our method at: https://github.com/mskimS2/GuideCAD

Recent post-training methods, particularly Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced the reasoning ability of Large Vision-Language Models (LVLMs). However, the sparse nature of verifiable rewards provides little token-level supervision for failed rollouts, often leading to inefficient exploration in complex multimodal reasoning tasks. Although policy distillation can offer dense guidance, external teacher based methods introduce substantial computational overhead, while answer conditioned tuning methods may expose answer-level information and induce shortcut-like generation behavior. To address these limitations, we propose PTD-PO, a Privileged Tutoring Distillation Policy Optimization framework for RLVR that provides dense guidance without exposing the answer to the student policy. Specifically, PTD-PO constructs structured privileged hints from spatial attention guidance and intermediate textual reasoning steps, and uses them through in-context learning to produce step-wise token-distribution supervision. The student is still optimized under the original answer-free context, and its failed rollouts are aligned with the hint-augmented reference model at the token-distribution level. To further stabilize distillation under the distribution shift between guided and unguided contexts, we introduce a Top-K Jensen-Shannon divergence objective that focuses alignment on informative token probabilities while reducing memory overhead. Experiments on LVLMs ranging from 2B to 8B parameters show that PTD-PO consistently outperforms RLVR and distillation baselines, mitigates entropy collapse, and improves complex multimodal reasoning performance.

Cross-domain shifts challenge Presentation Attack Detection (PAD) on ID Cards, given the restricted data available due to privacy concerns. This work proposes a compact multimodal model, based on new generative and discriminative blocks, which combines visual and textual data for PAD on genuine and synthetic ID images. While multimodal models exhibit strong generalisation after supervised fine-tuning, they fail in zero-shot settings. Our findings underscore that model capacity and real-world data are essential for reliable PAD, while existing synthetic datasets may not reflect real-world challenges. We argue for a re-evaluation of synthetic data as a benchmark and emphasise the need for more realistic, diverse datasets to advance PAD research.

Recent advancements have significantly expanded the theoretical context windows of Multimodal Embedding Models (MEMs). However, larger context windows do not necessarily translate into effective comprehension and representation of long-context multimodal inputs, which remains a critical bottleneck for real-world deployment. To address the lack of systematic evaluation in this setting, we introduce MMLongEmbed, the first comprehensive benchmark for evaluating MEMs in long-context scenarios. MMLongEmbed comprises four retrieval tasks spanning multiple context-length ranges, covering text, document, and video modalities. Through extensive evaluation of state-of-the-art models, we find that current architectures rely heavily on superficial feature matching and struggle to capture deep semantic and structural dependencies. We further observe that performance degradation varies systematically with context length and key information placement. Moreover, models exhibit substantially different robustness to redundant contextual information across modalities. For reproducibility, the benchmark and code are publicly available.

Vision and language models (VLMs) hold immense promise to transform biomedical imaging workflows, from detecting lesions in chest X-rays to profiling cellular features in microscopy. Realizing this potential, however, requires robust and fine-grained visual perception. Models need to correctly interpret subtle features in images, and they must do so across diverse biomedical modalities, scales, and contexts. Nevertheless, current benchmarks remain limited. To address these gaps, we introduce the Massive Multimodal Biomedical Understanding (MMBU) benchmark. It is the largest biomedical vision and language benchmark to date, covering 35 submodalities with rich structured metadata. It includes both open and closed versions of ungrounded classification, grounded classification, and object detection, enabling systematic evaluation of model performance across biological scales, clinical settings, and imaging modalities. Evaluating 15 open-weight and 2 frontier VLMs, we find that while medical adaptation provides measurable gains for some models, the high accuracy often reported on established benchmarks can mask deficiencies in visual perception and domain generalization.

Spatial transcriptomics (ST) is a powerful tool for exploring biological properties dependent on structure, proximity, and interaction in tissue. The methods underpinning ST are developing rapidly but are limited in their ability to profile many thousands of genes at a subcellular scale. Although dissociated from tissue, it is known that the whole-transcriptome readouts of cells in single-cell RNA sequencing (scRNA-seq) retain information about their former in situ neighbourhoods, motivating computational methods to recover it. While paired ST and scRNA-seq datasets are scarce, each modality in its own right is abundantly available. We therefore propose to perform cross-modal translation between unpaired ST and scRNA-seq data. In this work we show that a single-cell foundation model can perform this translation via adversarial fine-tuning. We demonstrate that our method performs favourably against methods built for multi-omics translation.

Recent advances in 3D multimodal large language models (3D-MLLMs) have enabled unified solutions for 3D scene understanding tasks, including visual question answering, captioning, and referring segmentation. However, existing 3D-MLLMs remain largely object-centric, limiting their ability to model fine-grained part structures that are essential for embodied interaction with 3D environments. In this work, we present PAR3D, a unified part-aware 3D-MLLM framework that enables models to understand, reason about, and ground both objects and their parts in 3D scenes. To enable training and evaluation of part-aware 3D scene understanding, we introduce ScenePart, a synthetic 3D scene dataset with part-level annotations and language instructions. We further develop Part-Aware 3D Representation Learning to enrich 3D visual representations with fine-grained part-level semantics, and propose Hierarchical Segmentation Query Generation to ground part targets via hierarchical object-part queries. Extensive experiments show that our method substantially improves part-level question answering and referring segmentation, while also achieving strong performance across object-level vision-language tasks.

Multimodal generative models produce fluent outputs but remain unreliable when generation must respect structured, domain-specific, or safety-critical knowledge. Existing methods incorporate knowledge through mechanisms such as prompt augmentation, guidance, latent editing, or fine-tuning, yet they are typically categorized by technique rather than by the component of the generative process they modify. We argue that knowledge infusion in iterative generative models is fundamentally anintervention-layer problem. Since thegenerative process unfolds as a trajectory of internal states, knowledge can act on four structurally distinct components of this process: the input/output boundary, the transition function, the intermediate state, and the model parameters. This maps to four intervention layers: surface, trajectory, latent, and parametric infusion. We instantiate the framework in diffusion models, map representative methods to all four layers, and derive design principles for multi-layer composition. In a controlled safety-alignment experiment using a multimodal knowledge graph with two diffusion backbones, we implement three of the four layers cumulatively, surface (input-side and output-side) and trajectory--latent (mid-generation). We show empirically that each additional layer addresses failure classes that prior layers cannot reach, reducing knowledge-violating outputs by 70.97% compared to vanilla generation and empirically confirming the framework's complementarity prediction.

Time series foundation models (TS-FMs) aim to learn generalizable temporal representations that can be adapted to a wide range of downstream tasks. In real-world multimodal settings, time series are frequently affected by temporal misalignment and partial modality missingness, where different modalities are observed at heterogeneous time scales or are partially absent. Existing approaches typically rely on naive imputation or masking strategies, which fail to account for cross-modal dependencies and often lead to misaligned or degraded representations. We propose TRACE, a conditional estimation paradigm for multimodal time series foundation model pipelines under missingness and irregular sampling, allowing incomplete target modalities to be systematically inferred from available auxiliary modalities. We evaluate TRACE on diverse multimodal benchmarks spanning healthcare and affective computing, including the MIMIC-IV clinical dataset and the CMU-MOSI and CMU-MOSEI benchmarks for multimodal sentiment analysis. Across a range of downstream prediction tasks and missing-modality settings, TRACE consistently outperforms prior multimodal fusion approaches, demonstrating improved robustness to severe modality missingness and more reliable cross-modal representations.

Transformer-based multimodal models rely on attention mechanisms to integrate information across heterogeneous modalities. Despite their success, existing multimodal attention formulations compute their scores through collections of pairwise dot-product interactions or by concatenating all the modalities into the keys, even when multiple modalities should be jointly involved. As a consequence, current approaches either incur quadratic complexity in the number of modalities or fail to explicitly model interactions that depend on the joint configuration of multiple representations. In this work, we introduce the Volumetric Multimodal cross-Attention (VMA), a novel cross-attention mechanism in which attention scores are defined as a function of the joint geometry of a query and multiple modality-specific keys. VMA computes the volume spanned by query and key vectors across multiple modalities, capturing joint multimodal dependencies beyond pairwise similarity, enabling native modeling of any-order modality interactions. We integrate VMA into our novel multimodal transformer architecture, named GRAMformer, explicitly designed to integrate any number of modalities. We evaluate the proposed model on multimodal learning tasks, demonstrating improved effectiveness and efficiency.

When a disaster unfolds, responders must answer not only what is happening, but also why it is happening, what will happen next, and what to do now, often from noisy low-altitude UAV views and under tight on-site compute constraints. However, most existing multimodal benchmarks emphasize perception (e.g., recognition/description), cover limited disaster types, and provide insufficient support for the multi-stage reasoning required in practical emergency response. We introduce DisasterBench, a multi-stage multimodal reasoning benchmark for UAV-Based disaster response in complex environments. DisasterBench spans 14 disaster-related scene types and 9 response-critical tasks across pre-, during-, and post-disaster stages, with fine-grained disaster-task mappings that explicitly test causal attribution, propagation prediction, damage analysis, and decision-oriented reasoning. To enable reasoning on the edge, we further propose DisasterVL, a lightweight multimodal model optimized with a three-stage pipeline combining domain instruction tuning, chain-of-thought-guided multimodal alignment, and reinforcement learning-based policy optimization. Experiments across 21 popular MLLMs show that our 2B-parameter DisasterVL outperforms all evaluated open-source models and substantially narrows the gap to state-of-the-art closed-source models, achieving GPT-4o-comparable reasoning accuracy with superior efficiency. The project page is available at https://github.com/TanmouTT/DisasterBench.

Developing unified video generation and editing models capable of interpreting interleaved multimodal inputs is a promising yet challenging frontier field. Existing unified frameworks predominantly rely on massive models (typically 13B parameters or more) and incorporate source video conditions for editing by concatenating sequence tokens. This concatenation inevitably doubles the sequence length, quadrupling the computational complexity of the self-attention mechanism and introducing prohibitive overhead. To address these bottlenecks, we present LoomVideo, a highly efficient 5B-parameter unified architecture for both video generation and editing. LoomVideo replaces the standard text encoder with a Multimodal Large Language Model (MLLM) and employs Deepstack injection mechanism to align multi-layer MLLM features with the Diffusion Transformer (DiT). Crucially, we introduce a zero-overhead Scale-and-Add conditioning approach for video editing. By scaling and directly adding the clean source video latent to the noised target latent, this elegant design eliminates the need for token concatenation, drastically reducing computational cost while maintaining robust capabilities for complex, non-rigid edits. Furthermore, a Negative Temporal RoPE strategy is seamlessly integrated to handle multiple reference images. Extensive experiments demonstrate that our compact 5B model achieves state-of-the-art or highly competitive performance across comprehensive benchmarks, exhibiting exceptional superiority in e-commerce and fashion generation scenarios. Benefiting from the zero-overhead conditioning mechanism, LoomVideo achieves at least a 5.41x acceleration in inference speed compared to models of similar capabilities, paving the way for highly practical and efficient video foundation models.

We present the AILS-NTUA submission to the EXIST 2026 Lab at CLEF, addressing multimodal sexism identification and characterization in memes (Task 2) and short-form videos (Task 3). Our system follows a feature-engineered late-fusion pipeline built around gradient-boosted regression models and hierarchical post-processing. For memes, we combine visual, textual, demographic, biometric, and LLM-derived semantic indicators designed to capture high-level cues such as stereotyping, objectification, irony, and misogyny. For videos, we investigate the effect of feature selection, frame-based visual representations, OCR-based textual features, acoustic descriptors, and sensor-derived metadata. Development results show that focused LLM-derived semantic cues improve meme sexism identification, while video performance is highly sensitive to feature dimensionality and cross-modal noise. For videos, development results favor compact feature selection, but official test results show that this conclusion does not fully transfer to unseen data, where the unfiltered representation generalizes better. Overall, our findings highlight the usefulness of targeted semantic feature engineering for static memes and the need for more robust temporal modeling in noisy short-form video settings.

Text-guided image editing has advanced rapidly with diffusion models and unified multimodal foundation models. However, most existing methods remain confined to single-turn settings, overlooking the more realistic scenario of multi-turn in-context editing, where users iteratively refine an image through a sequence of instructions. In this setting, a model must follow each new instruction while preserving accumulated session-level constraints, challenged by two coupled failure modes: long-context dilution, where sparse textual constraints become difficult to recover from growing interleaved image-text histories, and state contamination, where earlier editing mistakes degrade subsequent generations. We introduce Edit-R2, a novel reinforcement learning post-training framework for unified multimodal models. Edit-R2 reconstructs the operative session intent, which effectively consolidates scattered historical constraints into an explicit reasoning trace before each editing turn. It further enables multi-turn RL over both reasoning and generation through a unified objective that jointly optimizes intent reconstruction generation in discrete text space and flow-matching image generation in continuous latent space, while a trajectory filtering mechanism suppresses corrupted rollouts to stabilize training under state contamination. To support systematic evaluation, we introduce MICE-Bench, a large-scale benchmark for multi-turn in-context editing with automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA) over accumulated session constraints. Experiments show that Edit-R2 substantially improves multi-turn in-context editing and achieves competitive performance compared against strong baselines.

When retrieving a person from a video archive by voice and face, should the system be multimodal or not? In real-world broadcast archives, unlike curated benchmarks, a target may be heard but unseen, seen but unheard, or both. Fusing scores from an absent modality injects noise, degrading precision below the best unimodal system. We propose a query-adaptive framework that detects active modalities via cross-modal score consistency: when both modalities are active, files retrieved by one also score highly on the other; this agreement breaks down when a modality is absent. Classifiers driven by these cross-modal features achieve 89% detection accuracy. On the BBC Rewind corpus (with over 12,000 broadcast videos) the adaptive system attains 94.2% P@1, outperforming speaker-only (82.9%), face-only (93.4%), and fixed fusion (90.0%), recovering 64% of the gap to an oracle with ground-truth modality labels (96.6%).

Current evaluations for Multimodal Large Language Models (MLLMs) overwhelmingly focus on utility-driven objectives, leaving model behavior under logic-neutral scenarios largely underexplored. Stochasticity is essential in scenarios where multiple actions are equally valid, such as recommending travel itineraries or daily schedules where multiple options have similar utility. In such settings, deterministic policies may lead to repetitive behaviors and reduced coverage of valid alternatives. To bridge this gap, we propose RandomBench, a benchmark designed to evaluate whether MLLMs can maintain distributionally neutral behavior when selecting among equivalent options. We further introduce three metrics, including RI, BCI, BII, to quantify entropy and distributional bias. Experiments reveal a pervasive phenomenon termed Stochastic Collapse, where MLLMs fail to maintain uniform randomness under explicit random instructions, with top-1 probabilities reaching 97% from the ideal one quarter baseline and RI dropping to 0.068 in Claude Sonnet 4.6. Extensive ablation studies further demonstrate that these deviations persist across languages and representation formats, highlighting the robustness of distributional collapse in logic-neutral decision settings.

While Multimodal Large Language Models (MLLMs) demonstrate remarkable proficiency on complex vision-language tasks, the mechanisms by which they extract query-relevant visual features from complex, noisy contexts remain opaque. In this paper, we present an in-depth interpretability study that uncovers a profound structural property within MLLMs: functional sparsity in cross-modal retrieval. Leveraging a token-level metric termed Retrieval Attention Mass (RAM), we identify and characterize a highly specialized subset of attention heads, referred to as Context-aware Retrieval (CoRe) heads. Across diverse visual domains and model scales, we observe a clear functional division: CoRe heads act as dedicated information extractors, while most other heads distribute attention over broader contextual regions. Causal interventions further demonstrate the necessity of these specialized heads. Ablating only the top 5% of CoRe heads causes significant degradation in multimodal reasoning performance, whereas ablating lower-ranked heads has minimal effect. Moreover, acceleration experiments validate the utility of CoRe heads, showing that leveraging this localized sparsity significantly accelerates inference while maintaining robust task performance. Our findings reveal a structural principle of functional sparsity within MLLMs, refining the current understanding of mechanistic interpretability and laying a theoretical foundation that can inspire future architecture design and model optimization.

Multimodal Large Language Models (MLLMs) excel at 2D semantic understanding but lack intrinsic 3D awareness, resulting in representations that fail to maintain geometric and spatial consistency across video frames. Given the scarcity of large-scale 3D data, we present GeoVR, a novel framework that learns geometric representations using purely 2D video sequences. This approach effectively restructures the semantic latent space within MLLMs to unlock spatial intelligence. Rather than employing superficial feature mixing, GeoVR reshapes the internal representations of the MLLM by distilling geometry knowledge from pre-trained 3D foundation models. This is accomplished through a multi-objective learning strategy driven by four complementary geometric targets: (1) estimating inter-frame camera poses to embed varying viewpoint dynamics, (2) regressing dense depth maps to anchor physical distances, (3) predicting a metric scale factor for real-world calibration, and (4) distilling multi-scale 3D features to align the intermediate feature space. Guided by these explicit physical and geometric constraints, the model's internal representations naturally develop strong 3D awareness. Extensive experiments on spatial reasoning benchmarks demonstrate that GeoVR achieves state-of-the-art performance, establishing a new paradigm for endowing foundation models with spatial intelligence.

Video event prediction (VEP) requires models to infer unobserved future states from partial video evidence. Existing video MLLMs usually verbalize intermediate future reasoning in text space: once visual evidence is verbalized, fine-grained motion, geometry, and interaction cues can be lost, leading to plausible but visually ungrounded hallucinations. We introduce Future-L1, an interleaved latent visual reasoning framework that lets an MLLM alternate between language tokens and continuous latent visual spans during autoregressive decoding. To train this capability, we construct Future-L1-50K by selecting examples where future visual hints help prediction and align latent states to future-frame embeddings, then further optimize sampled latent trajectories with LA-DAPO, a latent-aware RL objective with outcome-contrastive and temporal-diversity rewards. Future-L1 achieves new state-of-the-art results on both benchmarks: on FutureBench, it improves Qwen3-VL-8B from 61.0 to 85.4 and exceeds the previous best Video-CoE by 10.4 points; on TwiFF-Bench, it improves the average score from 2.44 to 3.04. These results suggest that future-oriented video reasoning benefits from preserving intermediate visual semantics in latent space rather than translating every reasoning step into text.

Audio-Visual Speech Recognition (AVSR) enhances speech recognition robustness by leveraging visual cues, while real-world scenarios remain challenging due to viewpoint variation, audio distortion, and visual occlusion, which degrade modality quality and increase audio-visual asynchrony. In this paper, we propose a novel Modality-aware Multi-view Self-supervised representation framework for robust Audio-Visual Speech Recognition (M2S-AVSR). First, we introduce a multi-view representation learning encoder to learn view-invariant visual speech representations. Next, we employ a modality-aware module that explicitly models modality quality and cross-modal synchrony to perform fine-grained modality-aware fusion, enabling fine-grained visual information injection during decoding. In addition, we release AISHELL8-RealScene, a public multi-scenario, multi-view conversational audio-visual dataset recorded in real-world environments, and establish a speech recognition benchmark on it. Experiments on English and Mandarin benchmarks demonstrate the effectiveness of the proposed method under challenging conditions. On LRS3, M2S-AVSR achieves up to 29.4% relative improvement under viewpoint perturbation and visual degradation settings. Our method also achieves new state-of-the-art performance on the MISP2021-AVSR test set. On AISHELL8-RealScene, it achieves the best result in outdoor scenes. The proposed method and dataset provide useful support for future research on robust speech and multimodal tasks under realistic conditions.

Spatial planning maps are central to territorial governance, translating planning objectives, regulations, and spatial strategies into visual forms for decision-making, public communication, and institutional coordination. Their interpretation, however, requires fine-grained visual perception, spatial reasoning, and policy-informed professional judgment, creating major challenges for both human learners and AI systems. With the rapid progress of Vision-Language Models (VLMs), their use in urban planning analysis is gaining attention, yet existing multimodal benchmarks mainly target general visual understanding and overlook the domain-specific cognitive processes of planning practice. To address this gap, we introduce PlanBench-V, the first comprehensive benchmark for evaluating VLMs in spatial planning map interpretation. We first build the Spatial Planning Map Database (SPMD), an expert-annotated dataset of 223 planning maps and 1629 question-answer pairs curated by professional planners, covering diverse geographic regions and cartographic styles. We then propose a theory-informed evaluation framework assessing four progressive capabilities: Perception, Reasoning, Association, and Implementation, corresponding to the cognitive pipeline of planning map interpretation. Extensive experiments across two generations of VLMs show clear progress but persistent limitations. The best 2026 agentic reasoning model, Qwen3.6-Plus, substantially outperforms the best 2025 model, GPT-4o, by 27%. Nevertheless, all models still struggle with implementation-oriented tasks requiring evaluative judgment, policy sensitivity, and constraint-aware decision-making. These findings reveal fundamental limitations of current VLMs in professional planning contexts and highlight the need for domain-adaptive multimodal reasoning frameworks. Code and data are available at https://plangpt.github.io.

On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher's supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visually grounded reasoning. We propose ViCuR, a visually grounded privileged-teacher distillation framework that replaces answer-side privilege with visual cues (query-related evidence in the input). Because these cues are derived from the same visual input available at inference, their evidence is recoverable by the student. To support this, ViCuR introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary cue-generation losses. Across seven benchmarks with Qwen3-VL-2B and 8B students, ViCuR consistently improves over answer-based on-policy self-distillation by +1.19 and +1.24 on overall average performance. It also extends naturally to stronger-teacher OPD, surpassing OPD baselines by +0.64 and +1.08, with consistent out-of-domain gains at the 8B scale. These results show that, in multimodal on-policy distillation, the design of teacher privilege is as important as teacher strength.