Modality Gap

Multimodal Large Language Models (MLLMs) have been increasingly used as automatic evaluators-a paradigm known as MLLM-as-a-Judge. However, their reliability and vulnerabilities to biases remain underexplored. We find that many MLLM judges fail to reliably integrate key visual or textual cues, yielding unreliable evaluations when evidence is missing or mismatched, and exhibiting instability under semantically irrelevant perturbations. To address this, we systematically define Compositional Bias in MLLM-as-a-Judge systems and introduce MM-JudgeBias, a benchmark for evaluating it. MM-JudgeBias introduces controlled perturbations across Query, Image, and Response, and evaluates model behavior via two complementary metrics: Bias-Deviation (BD) for sensitivity and Bias-Conformity (BC) for stability. Our dataset of over 1,800 curated and refined multimodal samples, drawn from 29 source benchmarks, enables a fine-grained diagnosis of nine bias types across diverse tasks and domains. Experiments on 26 state-of-the-art MLLMs reveal systematic modality neglect and asymmetric evaluation tendencies, underscoring the need for more reliable judges.

Electroencephalography (EEG) foundation models have shown strong potential for learning generalizable representations from large-scale neural data, yet their clinical deployment is hindered by distribution shifts across clinical settings, devices, and populations. Test-time adaptation (TTA) offers a promising solution by enabling models to adapt to unlabeled target data during inference without access to source data, a valuable property in healthcare settings constrained by privacy regulations and limited labeled data. However, its effectiveness for EEG remains largely underexplored. In this work, we introduce NeuroAdapt-Bench, a systematic benchmark for evaluating test-time adaptation methods on EEG foundation models under realistic distribution shifts. We evaluate representative TTA approaches from other domains across multiple pretrained foundation models, diverse downstream tasks, and heterogeneous datasets spanning in-distribution, out-of-distribution, and extreme modality shifts (e.g., Ear-EEG). Our results show that standard TTA methods yield inconsistent gains and often degrade performance, with gradient-based approaches particularly prone to heavy degradation. In contrast, optimization-free methods demonstrate greater stability and more reliable improvements. These findings highlight the limitations of existing TTA techniques in EEG, provide guidance for future development, and underscore the need for domain-specific adaptation strategies.

Native Omni-modal Large Language Models (OLLMs) have shifted from pipeline architectures to unified representation spaces. However, this native integration gives rise to a critical yet underexplored phenomenon: modality preference. To bridge this gap, we first systematically quantify modality preference of OLLMs using a newly-curated conflict-based benchmark and the modality selection rate metric. Our evaluation of ten representative OLLMs reveals a notable paradigm shift: unlike the ``text-dominance'' of traditional VLMs, most OLLMs exhibit a pronounced visual preference. To further understand the underlying mechanism, we conduct layer-wise probing and demonstrate that such modality preference is not static but emerges progressively in the mid-to-late layers. Building upon these insights, we leverage these internal signals to diagnose cross-modal hallucinations, achieving competitive performance across three downstream multi-modal benchmarks without task-specific data. Our work provides both a mechanistic understanding and a practical tool for building more trustworthy OLLMs. Our code and related resources are publicly available at: https://github.com/icip-cas/OmniPreference

Domain generalization (DG) aims to maintain performance under domain shift, which in computer vision appears primarily as stylistic variations that cause models to overfit to domain-specific appearance cues rather than class semantics. To overcome this, recent methods use textual representations as stable, domain-invariant anchors. However, multimodal approaches that rely on cosine similarity-based contrastive alignment leave a modality gap where image and text embeddings remain geometrically separated despite semantic correspondence. We propose CrossFlowDG, a novel DG framework that addresses this residual gap using noise-free, cross-modal flow matching. By learning a continuous transformation in the joint Euclidean latent space, our framework explicitly transports domain-biased image embeddings toward domain-invariant text embeddings of the correct class. Using the efficient VMamba image encoder and CLIP's text encoder, CrossFlowDG is tested against four common DG benchmarks, and achieves competitive performance on several benchmarks and state-of-the-art on TerraIncognita. Code is available at: https://github.com/ajkrit/CrossFlowDG

Vision Language models (VLMs) have demonstrated strong performance across a wide range of benchmarks, yet they often suffer from modality dominance, where predictions rely disproportionately on a single modality. Prior approaches primarily address this issue by steering model's attention allocation, implicitly assuming that all modalities provide sufficient information. However, attention only determines where the model focuses, and cannot enrich information that is missing or ambiguous. In the real world, input modalities often differ in information density and their signal-to-noise ratios. In such cases, simply adjusting model's attention does not resolve the underlying lack of information. In this paper, we propose \textsc{MoIR}: \textit{Multi-modal Information Router}, an information-level fusion method that explicitly reduces information disparity prior to fusion. \textsc{MoIR} identifies less informative tokens and routes complementary information from a stronger modality, constructing information-dense token representations before they are processed by a large language model. By modifying information availability, \textsc{MoIR} enables reliable shifts in modality dominance, even when one modality is degraded. We evaluate \textsc{MoIR} on three widely used multi-modal benchmarks across multiple model backbones. Experimental results show that \textsc{MoIR} consistently demonstrates more balanced modality contribution, and improves robustness and downstream performance, particularly even under modality degradation. These findings demonstrate that explicitly modifying cross-modal information is an effective and complementary strategy for mitigating modality dominance in multi-modal reasoning models.

Reasoning in vision-language models (VLMs) has recently attracted significant attention due to its broad applicability across diverse downstream tasks. However, it remains unclear whether the superior performance of VLMs stems from genuine vision-grounded reasoning or relies predominantly on the reasoning capabilities of their textual backbones. To systematically measure this, we introduce CrossMath, a novel multimodal reasoning benchmark designed for controlled cross-modal comparisons. Specifically, we construct each problem in text-only, image-only, and image+text formats guaranteeing identical task-relevant information, verified by human annotators. This rigorous alignment effectively isolates modality-specific reasoning differences while eliminating confounding factors such as information mismatch. Extensive evaluation of state-of-the-art VLMs reveals a consistent phenomenon: a substantial performance gap between textual and visual reasoning. Notably, VLMs excel with text-only inputs, whereas incorporating visual data (image+text) frequently degrades performance compared to the text-only baseline. These findings indicate that current VLMs conduct reasoning primarily in the textual space, with limited genuine reliance on visual evidence. To mitigate this limitation, we curate a CrossMath training set for VLM fine-tuning. Empirical evaluations demonstrate that fine-tuning on this training set significantly boosts reasoning performance across all individual and joint modalities, while yielding robust gains on two general visual reasoning tasks. Source code is available at https://github.com/xuyige/CrossMath.

Fine-grained image retrieval via hand-drawn sketches or textual descriptions remains a critical challenge due to inherent modality gaps. While hand-drawn sketches capture complex structural contours, they lack color and texture, which text effectively provides despite omitting spatial contours. Motivated by the complementary nature of these modalities, we propose the Sketch and Text Based Image Retrieval (STBIR) framework. By synergizing the rich color and texture cues from text with the structural outlines provided by sketches, STBIR achieves superior fine-grained retrieval performance. First, a curriculum learning driven robustness enhancement module is proposed to enhance the model's robustness when handling queries of varying quality. Second, we introduce a category-knowledge-based feature space optimization module, thereby significantly boosting the model's representational power. Finally, we design a multi-stage cross-modal feature alignment mechanism to effectively mitigate the challenges of cross modal feature alignment. Furthermore, we curate the fine-grained STBIR benchmark dataset to rigorously validate the efficacy of our proposed framework and to provide data support as a reference for subsequent related research. Extensive experiments demonstrate that the proposed STBIR framework significantly outperforms state of the art methods.

Conventional frame-based cameras capture rich contextual information but suffer from limited temporal resolution and motion blur in dynamic scenes. Event cameras offer an alternative visual representation with higher dynamic range free from such limitations. The complementary characteristics of the two modalities make event-frame asymmetric stereo promising for reliable 3D perception under fast motion and challenging illumination. However, the modality gap often leads to marginalization of domain-specific cues essential for cross-modal stereo matching. In this paper, we introduce Bi-CMPStereo, a novel bidirectional cross-modal prompting framework that fully exploits semantic and structural features from both domains for robust matching. Our approach learns finely aligned stereo representations within a target canonical space and integrates complementary representations by projecting each modality into both event and frame domains. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods in accuracy and generalization.

Context-aided forecasting (CAF) holds promise for integrating domain knowledge and forward-looking information, enabling AI systems to surpass traditional statistical methods. However, recent empirical studies reveal a puzzling gap: multimodal models often fail to outperform their unimodal counterparts. We hypothesize that this underperformance stems from poor context quality in existing datasets, as verification is challenging. To address these limitations, we introduce a semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories. This approach enables massive-scale dataset creation, resulting in CAF-7M, a corpus of 7 million context-augmented time series windows, including a rigorously verified test set. We demonstrate that semi-synthetic pre-training transfers effectively to real-world evaluation, and show clear evidence of context utilization. Our results suggest that dataset quality, rather than architectural limitations, has been the primary bottleneck in context-aided forecasting.

The core of video-based visible-infrared person re-identification (VVI-ReID) lies in learning sequence-level modal-invariant representations across different modalities. Recent research tends to use modality-shared language prompts generated by CLIP to guide the learning of modal-invariant representations. Despite achieving optimal performance, such methods still face limitations in efficient spatial-temporal modeling, sufficient cross-modal interaction, and explicit modality-level loss guidance. To address these issues, we propose the language-driven sequence-level modal-invariant representation learning (LSMRL) method, which includes spatial-temporal feature learning (STFL) module, semantic diffusion (SD) module and cross-modal interaction (CMI) module. To enable parameter- and computation-efficient spatial-temporal modeling, the STFL module is built upon CLIP with minimal modifications. To achieve sufficient cross-modal interaction and enhance the learning of modal-invariant features, the SD module is proposed to diffuse modality-shared language prompts into visible and infrared features to establish preliminary modal consistency. The CMI module is further developed to leverage bidirectional cross-modal self-attention to eliminate residual modality gaps and refine modal-invariant representations. To explicitly enhance the learning of modal-invariant representations, two modality-level losses are introduced to improve the features' discriminative ability and their generalization to unseen categories. Extensive experiments on large-scale VVI-ReID datasets demonstrate the superiority of LSMRL over AOTA methods.

Recent vision-language models (VLMs) like CLIP have shown impressive anomaly detection performance under significant distribution shift by utilizing high-level semantic information through text prompts. However, these models often overlook fine-grained defect cues, e.g., hole, cut, or scratch, that are essential for understanding the anomaly's nature. Moreover, the modality gap between images and text can lead to subtle visual evidence being poorly captured in textual descriptions. To address the gap, we enhance the representation of "abnormal" with structured semantics, bridging coarse anomaly signals and fine-grained defect categories. We propose a hybrid prompting mechanism that combines human-readable descriptions of defect types with learnable token embeddings. Building on these ideas, we introduce DAPO, a Defect-aware Prompt Optimization framework for zero-shot multi-type and binary anomaly detection and segmentation under distribution shift. DAPO aligns anomaly-relevant visual features with their corresponding textual semantics by learning hybrid defect-aware prompts that combine fixed textual anchors with trainable token embeddings. We conducted experiments on public benchmarks (MPDD, VisA, MVTec-AD, MAD, and Real-IAD) and an internal dataset. The results suggest that compared to the baseline models, DAPO achieves a 3.6% average improvement in AUROC and average precision metrics at the image level under distribution shift, and a 5.2% average improvement in AUROC and F1 when localizing novel anomaly types under zero-shot settings.

Face Anti-Spoofing (FAS) is essential for the security of facial recognition systems in diverse scenarios such as payment processing and surveillance. Current multimodal FAS methods often struggle with effective generalization, mainly due to modality-specific biases and domain shifts. To address these challenges, we introduce the \textbf{M}ulti\textbf{m}odal \textbf{D}enoising and \textbf{A}lignment (\textbf{MMDA}) framework. By leveraging the zero-shot generalization capability of CLIP, the MMDA framework effectively suppresses noise in multimodal data through denoising and alignment mechanisms, thereby significantly enhancing the generalization performance of cross-modal alignment. The \textbf{M}odality-\textbf{D}omain Joint \textbf{D}ifferential \textbf{A}ttention (\textbf{MD2A}) module in MMDA concurrently mitigates the impacts of domain and modality noise by refining the attention mechanism based on extracted common noise features. Furthermore, the \textbf{R}epresentation \textbf{S}pace \textbf{S}oft (\textbf{RS2}) Alignment strategy utilizes the pre-trained CLIP model to align multi-domain multimodal data into a generalized representation space in a flexible manner, preserving intricate representations and enhancing the model's adaptability to various unseen conditions. We also design a \textbf{U}-shaped \textbf{D}ual \textbf{S}pace \textbf{A}daptation (\textbf{U-DSA}) module to enhance the adaptability of representations while maintaining generalization performance. These improvements not only enhance the framework's generalization capabilities but also boost its ability to represent complex representations. Our experimental results on four benchmark datasets under different evaluation protocols demonstrate that the MMDA framework outperforms existing state-of-the-art methods in terms of cross-domain generalization and multimodal detection accuracy. The code will be released soon.

Contrastive Language-Image Pre-training (CLIP) demonstrates strong zero-shot generalization, but due to substantial computational and memory costs, distillation into lightweight models is required. Existing relational objectives do not explicitly model multidirectional relationships between teacher and student embeddings, potentially leaving the geometric relationships insufficiently constrained. This may disrupt the modality-gap structure important for zero-shot transfer. To address these limitations, we propose a relational distillation framework, CLIP-RD, which introduces two relational methods, Vertical Relational Distillation (VRD) and Cross Relational Distillation (XRD). VRD aligns teacher-student intra-modal similarity distributions to enforce consistent distillation strength across image and text embeddings. Meanwhile, XRD aligns the teacher-image-student-text and teacher-text-student-image similarity distributions to impose bidirectional cross-modal symmetry. By jointly modeling these multidirectional relational structures, CLIP-RD aligns the student's embedding geometry more faithfully to the teacher's, outperforming CLIP-KD by 1.8%p. This performance improvement is maintained across diverse architectures, teacher scales, retrieval tasks, downstream tasks, and corruption settings, with negligible additional training-time overhead.