Cross-Modal Learning

Dataset distillation compresses large training sets into compact synthetic datasets while preserving downstream performance. As modern systems increasingly operate on paired vision-language inputs, multimodal distillation must preserve representation quality and cross-modal alignment under tight compute and memory budgets, yet prior methods often require heavy computes and overlook their correlations. To address this, we present Multimodal Distribution Matching (MDM), a geometry-aware framework for efficient and generalizable multimodal distillation. Specifically, MDM integrates complementary components at the data, model, and loss levels. At the data level, it initializes synthetic image-text pairs by sampling from clusters in the joint embedding space. At the model level, it forms a mixed teacher by interpolating independently fine-tuned models in weight space according to their angular deviation from the pretrained anchor. At the loss level, it matches joint distributions on the unit hypersphere using a geometry-aware matching objective that exploits the joint features in the cross-modal agreement and discrepancy directions along with symmetric contrastive learning. Across image-text retrieval benchmarks with cross-architecture evaluation, MDM yields compact synthetic sets that preserve multimodal semantics, substantially reduce distillation cost, and remain robust across architectures.

Multimodal learning often suffers from modality imbalance, where modalities that converge faster dominate optimization while others remain undertrained. Existing approaches typically mitigate this issue by strengthening the weak modality or adjusting optimization gradients. However, such strategies mainly compensate for optimization rate discrepancies, often at the expense of the strong modality's optimization capacity, without analyzing how these discrepancies arise at the modality level. Based on theoretical insights and empirical observations, we argue that the discrepancy of learning pace arises from differences in the mapping difficulty between modality-specific feature space and the shared label space. To address this issue, we propose Balanced Multimodal Label Reshaping (BMLR), the first method that promotes multimodal balance from the label-side design. BMLR reshapes the cross-modal label space to equalize mapping difficulty across modalities, thereby facilitating modality interaction and injecting richer inter-class information into each modality. Extensive experiments across multiple architectures demonstrate that BMLR consistently improves multimodal performance and exhibits strong compatibility with diverse model designs. The source code will be released soon.

Audio-visual deepfake localization demands interval-level outputs that serve as temporal evidence. Despite recent progress, symmetric fusion under single-sided or asynchronous forgeries propagates cross-modal noise, degrading high-precision localization. We present IaMSB, an inconsistency-aware multimodal Schrödinger Bridge (SB) that jointly estimates cross-modal consistency and performs interval-level localization. Unlike diffusion models, SB minimizes path-distribution discrepancy and yields consistency scores without explicit noise injection or denoising. With the Schrödinger Bridge (SB), IaMSB unifies consistency estimation, cross-modal information selection, and bridge-step scheduling in one framework. Specifically, a lightweight coarse bridge first proposes candidate intervals and estimates cross-modal consistency; these statistics select cross-modal witness signals and allocate bridge steps asymmetrically across modalities. A refinement bridge then performs step-tuned fusion and outputs refined, time-aligned intervals. IaMSB anticipates single-sided and asynchronous forgeries and, using bottlenecked cross-modal interaction with step allocation, suppresses noise transfer, avoids unnecessary iterations. Across benchmarks, IaMSB stabilizes strict-IoU boundary precision, raising [email protected] by 3%~10%, and yields improved high-precision localization, particularly for single-sided forgeries.

Vision-Language Models (VLMs) excel at tasks like zero-shot classification and cross-modal retrieval by mapping images and text to a shared space, but this requires expensive end-to-end training with massive paired datasets. Current post-hoc alignment methods reduce computational costs by connecting pretrained encoders through lightweight mappings, yet still demand substantial paired data. In this work, we investigate the potential of repurposing the classification heads of pretrained vision models as semantic prototypes. The recycling of these weights, typically discarded after pretraining, unlocks two distinct capabilities: it enables zero-shot alignment by using weights as semantic anchors, and serves as a robust data augmentation strategy by mixing these prototypes with real image-text pairs. We demonstrate that integrating our approach with several state-of-the-art post-hoc alignment techniques consistently boosts accuracy in cross-modal retrieval, zero- and few-shot classification tasks.

Joint audio-visual reasoning is essential for omnimodal understanding, yet current multimodal large language models (MLLMs) still struggle when reasoning requires fine-grained evidence from both modalities. A central limitation is that explicit text-based chain-of-thought (CoT) compresses continuous audio-visual signals into discrete tokens, weakening temporal grounding and shifting intermediate reasoning toward language priors. We argue that a unified latent space is a better medium for such reasoning because it preserves dense sensory information while remaining compatible with autoregressive generation. Based on this insight, we propose \textbf{LatentOmni}, a cross-modal reasoning framework that interleaves textual reasoning with audio-visual latent states. LatentOmni introduces feature-level supervision to align latent reasoning states with task-relevant sensory features and uses Omni-Sync Position Embedding (OSPE) to maintain temporal consistency between latent audio and visual states. We further construct \textbf{LatentOmni-Instruct-35K}, a dataset of audio-visual interleaved reasoning trajectories for supervising latent-space reasoning. Comprehensive evaluation across multiple audio-visual reasoning benchmarks demonstrates that LatentOmni achieves the best performance among the evaluated open-source models and consistently outperforms the Explicit Text CoT baseline, supporting latent-space joint reasoning as a promising path toward stronger omnimodal understanding.

Training-free zero-shot composed image retrieval models are recently gaining increasing research interest due to their generalizability and flexibility in unseen multimodal retrieval. Recent LLM-based advances focus on generating the expected target caption by exploring the compositional ability behind the LLMs. Although efficient, we find that 1) the generated captions tend to introduce unexpected features from the reference image due to the semantic gap between the input image and text modification, where the image contains much more details than the text; 2) the point-to-point alignment during the retrieval stage fails to capture diverse compositions. To address these challenges, we introduce a novel Semantic Transition and Transportation in collaboration framework for training-free zero-shot CIR tasks. Specifically, given the composed caption inferred by an LLM, we aim to refine it through a transition vector in the embedding space and make it closer to the target image. Combining LLMs with user instruction, the refined caption concentrates more on the core modification intent and thus filters out unnecessary noise. Moreover, to explore diverse alignment during the retrieval stage, we model the caption and image as discrete distributions and reformulate the retrieval task as a set-to-set alignment task. Finally, a bidirectional transportation distance is developed to consider fine-grained alignments across modalities and calculate the retrieval score. Extensive experiments demonstrate that our method can be general, effective, and beneficial for many CIR tasks.

Despite the impressive results achieved by multimodal large language models (MLLMs), their training typically relies on jointly curated multimodal data, requiring substantial human effort to construct multi-way aligned datasets and thereby limiting scalability across domains. In this work, we explore training MLLMs by only leveraging multiple paired modalities as a surrogate for the full joint multimodal distribution. Specifically, we first provide a theoretical analysis of the conditions under which the representations are identifiable with only observing pairwise modalities. Building on this analysis, we propose a representation learning framework for aligning latent representations across modalities using only pairwise data. The framework consists of two stages: latent representation alignment and cross-modal recomposition. Specifically, in the first stage, we learn the shared latent space across modalities by both self-modal reconstruction and pair-wise contrastive learning. We also incorporate an inductive bias in the contrastive learning process by partially aligning and minimal latent specification. In stage two, we integrate the encoder of newly introduced modalities with the decoders of the pre-trained modalities to facilitate cross-modal transfer and generation. We evaluate our method by newly adding 3D point clouds and tactile modalities into pre-trained MLLMs with three modality pairs and show that, by learning an aligned latent representation space, our model achieves strong cross-modal performance.

Can language-pretrained transformers become effective time-series forecasters, and why? In this paper, we show that cross-modal transfer arises because language pretraining preconditions time series training with a reusable manifold. A linear probe on frozen LLM states decodes realistic time-series trajectories without paired supervision, and retrieval in this projected space yields competitive forecasts, showing that structure and dynamics exist before finetuning. Pretrained initialization also improves optimization, producing coherent gradients and a highly anisotropic loss landscape unlike random initialization. Finetuning then acts as low-dimensional alignment, reusing existing directions rather than learning temporal primitives from scratch, as evidenced by low-rank updates, subspace alignment, and shared features for periodicity, trend, and repetition. Together, these results support a geometric account of LLM-to-time-series transfer: language pretraining builds the manifold, and finetuning projects numerical dynamics onto task-relevant directions.

Multi-modal remote sensing images are vital for Earth observation, yet complete paired observations are often scarce in practice. Existing generative methods commonly address this problem through isolated pairwise modality translation, but their versatility and scalability remain limited as the number of modalities and generation tasks increases. Here, we develop a generative foundation model MetaEarth-MM for multi-modal remote sensing imagery, enabling paired joint generation and any-to-any translation across five modalities within a unified model. Recognizing the intrinsic scene consistency underlying multi-modal observations, we introduce a scene-centered joint modeling paradigm in MetaEarth-MM. Unlike previous methods that rely on direct appearance-level cross-modal mapping, our model organizes the generation around the underlying scene content. Specifically, MetaEarth-MM adopts a decoupled architecture that first infers a latent scene representation from available observations, and then generates target modalities conditioned on this intermediate state. To support training, we further construct EarthMM, a large-scale dataset comprising 2.8 million multi-resolution global images with 2.2 million aligned pairs. Extensive experiments demonstrate that MetaEarth-MM not only exhibits strong generative capability and robust generalization across diverse generation tasks, but also supports downstream tasks at both data and representation levels, highlighting its potential as a general foundation model for cross-modal Earth observation. The code and dataset will be available at https://github.com/YZPioneer/MetaEarth-MM.

Humans infer emotions by integrating observed multimodal cues with expectations about how affective states may unfold. Existing multimodal large language models (MLLMs), however, often treat emotion recognition as static fusion over complete audiovisual-text inputs, leaving affective dynamics implicit. We propose AffectVerse, a Qwen2.5-Omni-based model equipped with an Emotion World Module (EWM), an action-free representation-level module for short-horizon latent affective prediction. \rev{EWM contains three modules: 1) Cross-Modal Temporal Imagination predicts future video/audio representations from past tokens with multi-step rollout. 2) MAMA(Modality-Aware Multi-step Attention) Belief Aggregation compresses imagined tokens into modality-aware belief tokens. 3) Belief Injection inserts these belief tokens into the LLM for affective reasoning.} AffectVerse uses future prediction as a past-conditioned self-supervised signal: it does not replace modeling observed history or require unseen signals at inference, but forces the current belief state to encode transition cues that are predictive of subsequent affective change. Across nine benchmarks, AffectVerse improves at least 2.57% over other models, while controlled ablations show additive gains from temporal imagination, cross-modal rollout, and belief aggregation. These results suggest predictive belief-state modeling is a practical alternative for affective computing.

Neuroscientists have recently turned to intracranial brain recording methods, like electrocorticography (ECoG), for human experiments because of the fine spatial and temporal resolution that they afford. Models trained on this data, however, are fundamentally restricted by the patient populations that can receive the implants necessary for recording. We propose using non-invasive fMRI to bridge the gap in training data. Using spoken language representations fine-tuned on fMRI, we build encoding models of ECoG. These representations showed improved prediction performance in ECoG, even though the temporal resolution of fMRI is two orders of magnitude worse. Prediction improved in frequency bands well beyond what is directly measured in fMRI. Next, to test the procedure's generalization ability, we fine-tuned models on fMRI responses that were temporally downsampled by a factor of 2. Despite the loss in resolution, these models were able to predict fMRI and ECoG responses at levels comparable to the original fMRI-tuned models. Finally, we showed that ECoG performance steadily scales with the amount of fMRI-tuning data. Our results show that "slow" data like fMRI can be a valuable resource for building better models of "fast" brain data like ECoG. In the future, integrating across multiple recording methods may further improve performance in other applications, like decoding.

Real-world time series come with text: metadata, descriptions, news, reports. Yet time series foundation models process numerical sequences in isolation, and the multimodal text-and-time-series models that attempt to bridge the two all adapt a pretrained language model post hoc, inheriting representations shaped without ever seeing temporal data. These models are also evaluated almost exclusively against other multimodal baselines, not against the strongest unimodal foundation models in either domain, leaving open whether joint training is needed at all. We present Chronicle, a compact 324M-parameter decoder-only transformer trained from scratch on natural language and time series within a single unified architecture. Both modalities share the same transformer blocks, attention mechanism, and residual stream; the bulk of pretraining uses unimodal batches so cross-modal capability emerges purely from shared parameters, with a short alignment stage that interleaves the two. To our knowledge, Chronicle is the first model jointly pretrained on text and time series from scratch, and the first multimodal model evaluated against dedicated foundation models in both domains. It matches Gemma-3-270M-PT on 19 NLU tasks, sets a new bar for frozen-embedding time series classification on 24 UCR/UEA datasets, and produces multimodal forecasts on Time-MMD that beat every supervised fusion baseline, all from a single backbone.

Due to the difficulty of obtaining ground-truth data for 4D radar scene flow estimation, previous methods typically rely on either self-supervised losses or cross-modal supervision using 3D LiDAR data, 2D images, and odometry. However, self-supervised approaches often yield suboptimal results due to radar's inherently low-fidelity measurements, while existing cross-modal supervised methods introduce complex multi-task architecture and require costly LiDAR sensors to generate pseudo radar scene flow labels from pretrained 3D tracking models. To overcome these limitations, we propose a task-specific iterative framework for weakly supervised radar scene flow learning, using only images and odometry for auxiliary supervision during training. Specially, we establish two novel instance-aware self-supervised losses by exploiting off-the-shelf 2D tracking and segmentation algorithms to obtain tracked instance masks, which are back-projected into 3D space to provide instance-level semantic guidance; for static regions, we integrate vehicle odometry with radar's intrinsic motion cues to construct a rigid static loss. Extensive experiments on the real-world View-of-Delft (VoD) dataset demonstrate that our method not only surpasses state-of-the-art cross-modal supervised approaches that rely on 3D multi-object tracking on dense LiDAR point clouds but also outperforms existing fully supervised scene flow estimation methods. The code is open-sourced at https://github.com/FuJingyun/IterFlow.

Multimodal representation alignment is pivotal for large language models and robotics. Traditional methods are often hindered by cross-modal information discrepancies and data scarcity, leading to suboptimal alignment spaces that overlook modality-unique features. We propose CodeBind, a framework that optimizes multimodal representation spaces through a modality-shared-specific codebook design. By incrementally aligning target and bridging modalities, CodeBind bypasses the need for fully paired data. Unlike traditional hard alignment, CodeBind decomposes features into shared components for semantic consistency and specific components for modality-unique details. This design utilizes a compositional vector quantization scheme, where a shared codebook bridges modality gaps and modality-specific codebooks mitigate representation bias by preventing dominant modalities from overshadowing others. Validated across nine modalities (text, image, video, audio, depth, thermal, tactile, 3D point cloud, EEG), CodeBind achieves state-of-the-art performance in multimodal classification and retrieval tasks.

We present SWIM (See What I Mean), a novel training strategy that aligns vision and language representations to enable fine-grained object understanding solely from textual prompts. Unlike existing approaches that require explicit visual prompts, such as masks or points, SWIM leverages mask supervision only during training to guide cross-modal attention, allowing the model to automatically attend to the user-specified object at inference. Our cross-attention analysis of pretrained multimodal large languagemodels (MLLMs) reveals a systematic discrepancy: Attribute words produce sharp, localized activations in the visual modality, whereas object nouns yield diffuse and scattered patterns due to semantic reference bias and distributed high-level representations. To address this misalignment, we construct NL-Refer, an enriched dataset, in which each object mask is paired with a precise natural language referring expression. SWIM extracts multi-layer cross-attention maps from object nouns and enforces spatial consistency with ground-truth masks. Experimental results demonstrate that SWIM substantially improves text-visual alignment and achieves superior performance over visual-prompt-based methods on fine-grained object understanding benchmarks. The code and data are available at https://github.com/HumanMLLM/SWIM.

Unified multimodal models (UMMs) achieve strong performance in both understanding and generation by learning a shared latent space, yet they often exhibit functional inconsistency between these two capabilities. We observe that this issue does not stem from a lack of shared representations, but from the absence of explicit alignment between the transformations that map into and out of the latent space. As a result, generation and re-encoding can follow inconsistent trajectories, leading to semantic drift under modality transitions. In this work, we propose LatentUMM, a framework that constructs an enhanced shared latent space to explicitly align these transformations and improve cross-modal consistency. LatentUMM consists of two stages. First, dual latent alignment enforces consistency at both the modality and capacity levels: cross-modal alignment uses a stronger embedding model to impose structured cross-modal semantics, while dual capacity alignment enforces bidirectional consistency under generation and re-encoding. Second, latent dynamics stabilization improves robustness via stochastic latent rollouts and preference optimization, favoring trajectories that better preserve semantic consistency. Experiments show that LatentUMM consistently improves multimodal consistency across diverse architectures. Code is available at: https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/LatentUMM.

Multimodal deep neural networks enhance deep comprehension by integrating diverse data modalities. Data from different modalities are typically projected into a shared latent space for similarity computation, but this process is resource intensive due to large image encoders and equal processing of test data during prediction. Early exit methods reduce computational load by utilizing intermediate layers, saving time and memory. However, developing such methods is challenging for multimodal data like image-text pairs. This study investigates the semantic content distributions present in intermediate layers of encoders such as CLIP, which can be derived from textual descriptions. We introduce Text-Guided Exit Modules (T-GEMs) and a rate-based regularizer to control encoder usage costs while maintaining cross-modal understanding performance.

Cross-modality recognition has many important applications in science, law enforcement and entertainment. Popular methods to bridge the modality gap include reducing the distributional differences of representations of different modalities, learning indistinguishable representations or explicit modality transfer. The first two approaches suffer from the loss of discriminant information while removing the modality-specific variations. The third one heavily relies on the successful modality transfer, could face catastrophic performance drop when explicit modality transfers are not possible or difficult. To tackle this problem, we proposed a compact encoder-decoder neural module (cmUNet) to learn modality-agnostic representations while retaining identity-related information. This is achieved through cross-modality transformation and in-modality reconstruction, enhanced by an adversarial/perceptual loss which encourages indistinguishability of representations in the original sample space. For cross-modality matching, we propose MarrNet where cmUNet is connected to a standard feature extraction network which takes as inputs the modality-agnostic representations and outputs similarity scores for matching. We validated our method on five challenging tasks, namely Raman-infrared spectrum matching, cross-modality person re-identification and heterogeneous (photo-sketch, visible-near infrared and visible-thermal) face recognition, where MarrNet showed superior performance compared to state-of-the-art methods. Furthermore, it is observed that a cross-modality matching method could be biased to extract discriminant information from partial or even wrong regions, due to incompetence of dealing with modality gaps, which subsequently leads to poor generalization. We show that robustness to occlusions can be an indicator of whether a method can well bridge the modality gap.

Multimodal pre-training demonstrates strong generalization performance, but this paradigm is often impractical in domains where paired data are scarce. A promising alternative is post-hoc multimodal alignment, which aligns separately pre-trained unimodal encoders using a limited number of paired examples. However, existing methods focus primarily on aligning global representations, missing patch-token relations. This may hinder transfer to tasks that require fine-grained cross-modal matching beyond coarse sample-level semantics. To address this issue, we propose a post-hoc alignment method that learns token-level cross-modal structure using relative representations. Specifically, we represent images and texts through their token-level similarities to a set of learnable anchors in each modality space, which are trained to induce consistent cross-modal similarity patterns for matched pairs. Despite learning only the anchors without heavy projection layers, our approach consistently outperforms existing methods in zero-shot classification, cross-modal retrieval, and zero-shot segmentation by a substantial margin. This highlights the importance of modeling fine-grained cross-modal structure for effective post-hoc multimodal alignment with limited paired data.

Collaborative game-based learning environments offer rich opportunities for small-group knowledge construction, yet automatically predicting student collaboration satisfaction remains challenging. A critical barrier is modality degradation: in educational deployments, individual modalities such as eye gaze exhibit inconsistent informativeness across student cohorts, causing implicit attention-based fusion to produce brittle multimodal representations. We propose the Affinity-Aligned Multimodal Learning Analytics (AAMLA) framework, whose core contribution is the Cross-modal Affinity-guided Modality Alignment (CAMA) module, which explicitly models inter-modal relationships via affinity matrices and enforces cross-modal consistency through contrastive learning, enabling adaptive suppression of uninformative modalities without discarding them. AAMLA further applies modality-specific projection layers to map heterogeneous features, including facial action units, head pose, eye gaze, and interaction trace logs, into a unified semantic space prior to alignment. Experiments on 50 middle school students in the EcoJourneys collaborative learning environment demonstrate consistent improvements over unimodal baselines and prior cross-attention approaches under standard and modality degradation conditions, with SHAP and t-SNE analyses confirming that CAMA produces robust, interpretable cross-modal representations for student collaboration modeling.

In this work, we address the critical yet underexplored challenge of symmetric multimodal-to-multimodal (MM2MM) retrieval, where queries and contexts are interchangeable. Existing universal multimodal retrieval works struggle with this task, as they are constrained by the labeled asymmetric datasets used. We produce SOLAR (Self-supervised jOint LeArning for symmetric multimodal Retrieval), a novel two-stage self-supervised framework that leverages readily available unlabeled web-scale image-text pairs. Based on the observation that both semantic alignment and discrepancies exist between two modalities, in the first stage, we learn the intersection mask of image-text pair, allowing us to align intersection while preserving semantic of difference. In the second stage, the learned mask is further utilized to construct positive and hardnegative samples via masking different parts of image/text, which enable us to conduct self-supervised multimodal embedding learning. Complementing this framework, we present a new benchmark featuring high-quality human-verified positive and hard-negative pairs to evaluate symmetric MM2MM retrieval under realistic conditions, as well as the corresponding pipeline. Extensive experiments against ten SOTA methods show SOLAR surpasses the strongest supervised VLM by 7.08 points on this benchmark, with over 50x fewer model parameters and a 5x smaller embedding dimension. Code, model and benchmark are available at https://github.com/codefuse-ai/SOLAR.

Multimodal retrieval uses images, text, and object relationships to answer different questions about the same collection. A query may seek an object's paired description, another object in the same category, or an object connected by an observed relationship. These goals rely on different notions of relevance. Paired matching requires object-specific distinctions, whereas cross-object retrieval benefits from relational agreement. Existing methods learn strong cross-modal correspondence or propagate information over a graph, but a shared output regularized toward neighbors can weaken identity distinctions. Moreover, gains from graph regularization can diminish after uniform graph propagation. We introduce Graph-Optimized Multimodal Alignment (GOMA), which assigns these roles to two connected embeddings. Each modality produces a content embedding directly supervised for paired identity and a semantic embedding jointly trained with cross-modal pairs and observed relationships. For complete records, both embeddings form an initial fused representation, semantic agreement sets positive weights on observed edges, and restart graph propagation reinjects this initial signal. This design lets a jointly trained model support single-modality and dual-attribute retrieval through a task-specific readout. Across six datasets and four tasks, GOMA achieves state-of-the-art performance on all 14 primary measures against 14 external methods. Controlled comparisons further show how separate supervision, graph regularization, and semantic-guided graph propagation shape the final representation and align the learned signal with each retrieval target.

Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(ε,δ)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.

Standard dual-encoder vision-language models that map images and text to deterministic points on a shared unit hypersphere through $\ell_2$ normalization typically expose neither \emph{aleatoric} uncertainty (cross-modal ambiguity) nor \emph{epistemic} uncertainty (lack of training-distribution support). Existing post-hoc methods either recover at most one of the two uncertainty components, or ignore the hyperspherical geometry of these models' embeddings. We propose \textbf{GeoFlowVLM} as a post-hoc adapter that learns the joint distribution of paired $\ell_2$-normalised dual-encoder VLM embeddings on the product hypersphere $\mathbb{S}^{d-1} \times \mathbb{S}^{d-1}$ via Riemannian flow matching with a single masked velocity field. A consistency result shows that, in the population limit, the trained network exposes the joint flow and both cross-modal conditional flows as valid Riemannian flow-matching velocity fields on their respective domains. We derive two quantities from this single model: a conditional retrieval entropy that quantifies aleatoric ambiguity with a decision-theoretic interpretation via a Fano-type bound, and a marginal-typicality epistemic score justified by an exact chain-rule decomposition of the joint NLL. This decomposition isolates a cross-modal pointwise-mutual-information term that is structurally discriminative rather than epistemic, and is empirically the only consistently uninformative standalone component. Empirically, the entropy tracks Recall@1 with near-ideal monotonic calibration across three retrieval benchmarks in both directions, and the marginal-typicality sum yields consistently calibrated selective accuracy across four zero-shot classification benchmarks.

Despite the recent success of Multimodal Foundation Models (FMs), their reliance on massive paired datasets limits their applicability in low-data and rare-scenario settings where aligned data is scarce and expensive. A key bottleneck is the adoption of an instance-level formulation, which learns alignment by maximizing correlation between individual image-text pairs while neglecting the underlying geometric structure across modalities resulting in a modality gap across input modalities. In this paper, we propose a combinatorial paradigm for multimodal alignment that moves beyond pairwise learning and introduce the \emph{Submodular Modality Aligner (SMA)}, which treats multiple augmentations and descriptions of an entity as a set, leveraging multiple descriptions of the data to capture richer cross-modal structure. We instantiate SMA using a principled objective based on Submodular Mutual Information (SMI), which jointly maximizes inter-modality mutual information while reducing cross-modal divergence. This formulation enables the model to effectively utilize multiple positive associations and extract significantly more information from limited data. We evaluate SMA on 14 zero-shot classification and retrieval tasks from the CLIP benchmark and demonstrate consistent gains in the low-data regime. Notably, SMA achieves strong multimodal generalization using only tens of thousands of samples. This is orders of magnitude fewer than standard approaches. Our results highlight the importance of set-based formulations and submodular objectives for data-efficient multimodal learning.

A central objective in multimodal learning is to capture synergy: task-relevant information that arises only from the joint use of multiple modalities, and is not available from any single modality alone. While most approaches operate at the architectural level through larger or more complex fusion models, we propose a complementary axis: shaping the training objective itself. Standard training often emphasizes unimodal or redundant information, falling short on examples that require cross-modal reasoning. We formalize multimodal synergy through information theory and introduce the Synergistic Information Bottleneck (SynIB), a scalable objective that targets synergy directly. To prioritize learning synergy, SynIB motivates the model to predict accurately from all modalities while penalizing confidence when information from any modality is withheld. Alongside the standard task loss, the model runs forward passes with one modality masked at a time and is penalized for remaining confident, which would indicate reliance on unimodal cues rather than cross-modal interactions. We validate SynIB in two regimes. On synthetic XOR tasks where the ground-truth synergy is known by construction, standard training fails to recover it while SynIB does. On five real-world benchmarks, including three MultiBench affective tasks, Hateful Memes with CLIP-ViT and DeBERTa backbones, and a controllable irony extension of CREMA-D we introduce, SynIB improves accuracy on synergy-dependent examples by up to 7.8% and overall accuracy by up to 3.8%.

Multimodal learning seeks to integrate information across diverse sensory sources, yet current approaches struggle to balance cross-modal generalizability with modality-specific structure. Continuous (implicit) methods preserve fine-grained priors but render generalization challenging, while discrete (explicit) approaches enforce shared prototypes at the expense of modality specificity. We introduce CoDAAR (Cross-modal Discrete Alignment And Reconstruction), a novel framework that resolves this long-standing trade-off by establishing semantic consensus across modality-specific codebooks through index-level alignment. This design uniquely allows CoDAAR to preserve modality-unique structures while achieving generalizable cross-modal representations within a unified discrete space. CoDAAR combines two complementary mechanisms: Discrete Temporal Alignment (DTA), which enables fine-grained temporal quantization, and Cascading Semantic Alignment (CSA), which promotes progressive cross-modal semantic agreement. Together, they establish a competition-free unified representation space. Trained with self-supervised reconstruction objectives on paired multimodal sequences, CoDAAR demonstrates robust cross-modal and cross-domain generalization. Across Cross-Modal Generalization benchmarks, including event classification, localization, video segmentation, and cross-dataset transfer, CoDAAR achieves state-of-the-art performance, establishing a new paradigm for discrete and generalizable multimodal representation learning.

Omni-modal language models are intended to jointly understand audio, visual inputs, and language, but benchmark gains can be inflated when visual evidence alone is enough to answer a query. We study whether current omni-modal benchmarks separate visual shortcuts from genuine audio-visual-language evidence integration, and how post-training behaves under a visually debiased evaluation setting. We audit nine omni-modal benchmarks with visual-only probing, remove visually solvable queries, and retain full subsets when filtering is undefined or would make comparisons unstable. This yields OmniClean, a cleaned evaluation view with 8,551 retained queries from 16,968 audited queries. On OmniClean, we evaluate OmniBoost, a three-stage post-training recipe based on Qwen2.5-Omni-3B: mixed bi-modal SFT, mixed-modality RLVR, and SFT on self-distilled data. Balanced bi-modal SFT gives limited and uneven gains, RLVR provides the first broad improvement, and self-distillation reshapes the benchmark profile. After SFT on self-distilled data, the 3B model reaches performance comparable to, and in aggregate slightly above, Qwen3-Omni-30B-A3B-Instruct without using a stronger omni-modal teacher. These results show that omni-modal progress is easier to interpret when evaluation controls visual leakage, and that small omni-modal models can benefit from staged post-training with self-distilled omni-query supervision. Project page: https://cheliu-computation.github.io/omni/

Audio-visual large language models (AVLLMs) have recently emerged as a powerful architecture capable of jointly reasoning over audio, visual, and textual modalities. In AVLLMs, the bidirectional interaction between audio and video modalities introduces intricate processing dynamics, necessitating a deeper understanding of their internal mechanisms. However, unlike extensively studied text-only or large vision language models, the internal workings of AVLLMs remain largely unexplored. In this paper, we focus on cross-modal information flow between audio and visual modalities in AVLLMs, investigating where information derived from one modality is encoded within the token representations of the other modality. Through an analysis of multiple recent AVLLMs, we uncover two common findings. First, AVLLMs primarily encode integrated audio-visual information in sink tokens. Second, sink tokens do not uniformly hold cross-modal information. Instead, a distinct subset of sink tokens, which we term cross-modal sink tokens, specializes in storing such information. Based on these findings, we further propose a simple training-free hallucination mitigation method by encouraging reliance on integrated cross-modal information within cross-modal sink tokens. Our code is available at https://github.com/kaistmm/crossmodal-hub.

Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, yet real-world deployment often requires continual capability expansion across sequential tasks. In such scenarios, Multimodal Continual Instruction Tuning (MCIT) aims to acquire new capabilities while limiting catastrophic forgetting. Existing methods mainly follow a module-composition paradigm: they maintain task-level prompts or LoRA experts and dynamically route or aggregate a subset of them at inference. However, samples within the same task can still differ substantially in visual scenes, question intents, and reasoning demands. This motivates instance-level adaptation to individual query-image pairs rather than only selecting or combining task-level modules. To this end, we propose DRAPE (Dynamic Cross-Modal Prompt Generation), a prompt-learning framework that synthesizes continuous instance-specific soft prompts for MCIT. Instead of selecting prompts from a fixed pool, DRAPE derives prompt queries from the textual instruction and cross-attends to visual patch features, producing query-image conditioned prompts that are prepended to the frozen LLM. To mitigate forgetting during sequential updates, DRAPE applies null-space gradient projection to the shared projector and uses CLIP-based prototype routing for task-label-free generator selection at inference. Extensive experiments on MCIT benchmarks show that DRAPE achieves state-of-the-art performance among representative prompt-based and LoRA-based continual-learning baselines.