Cross-Modal Knowledge Distillation

Open-vocabulary detection (OVD) recognizes categories unseen during training through textual category queries, yet achieving strong generalization with real-time efficiency remains challenging. Beyond vocabulary scaling, zero-shot generalization may benefit from reusable visual--semantic cues learned from seen data, including attributes, actions, states, and contextual relations. Existing real-time OVD methods primarily emphasize vocabulary coverage and efficient region/query--text matching; under strict efficiency constraints, compact detectors may struggle to absorb rich instance semantics and scene context. We propose RT-DETR-World, a compact DETR-style detector that transfers the rich semantics conveyed by descriptions during training while retaining lightweight query--text matching at inference. We construct GroundingCapv2 with three levels of supervision: category names for standard OVD, object descriptions conveying instance-level semantics, and image descriptions conveying object relations and scene context. These descriptions serve only as training-time semantic supervision. To help the compact detector absorb these semantics, we propose Dual-Path Description Alignment (DDA), combining a deployment-consistent MiniLM pathway with a training-only LLM teacher. MiniLM provides query--category supervision and object-description alignment, while offline teacher features supervise matched queries and global visual representations at the object and image levels, respectively. All teacher features are precomputed, and the teacher-side modules are removed after training. We further propose Relation-Aware Negative Relaxation (RNR), which uses teacher-derived semantic similarities to relax related negatives while preserving exact positives. Experiments demonstrate competitive zero-shot accuracy and a favorable accuracy--efficiency trade-off. The code will be released.

World-action models (WAMs) such as Cosmos 3 jointly generate future video and robot actions from an observation and instruction. Adapting one such model with a lightweight LoRA fine-tune to a previously unseen robot, a Unitree G1 humanoid with five-fingered BrainCo hands, exposes a video-action asymmetry: the video renders plausible task executions, while the co-generated action is systematically mis-targeted. We evaluate closed-loop real-robot trials at three cumulative stages: pre-grasp, grasp, and pick-and-place. The native action succeeds only approximately 17%, 10%, and 7% of the time, respectively, and performs worse on held-out objects. We propose AutodidactWAM, a hand-pose estimator trained without teleoperation, followed by inverse kinematics, that runs on the model's generated video to recover action estimates. Paired with the native prediction, these recovered actions provide preferred targets for fine-tuning only the action-related layers, while the generated video is teacher-forced. We compare supervised relabeling with a rectified-flow adaptation of Diffusion-DPO. After one-time embodiment adaptation, self-distillation requires no additional task-specific teleoperation. The recovered-action gate reaches approximately 75%, 47%, and 42% pre-grasp, grasp, and pick-and-place success, compared with 17%, 10%, and 7% for the native action. A hybrid objective combining preference supervision, supervised target fitting, and Cartesian trajectory anchoring (DPO+SFT+DTW) performs best: on Oreo, the training object, it reaches 90% pre-grasp and 20% full-task success; on a held-out object, it reaches 80% and 30%. Plain Flow-DPO reaches 0% success despite 1.000 validation preference accuracy, indicating that the combination of training objectives, rather than the contrastive objective alone, drives the observed gains.

Gastric adenocarcinoma (GA) is a leading cause of cancer-related mortality worldwide, and accurate histopathological subtype classification from whole-slide images (WSIs) is essential for effective treatment planning. While multimodal approaches that integrate pathology report text with WSIs can improve classification, existing methods often depend on computationally expensive transformer architectures and large language models. We propose a multimodal knowledge distillation (MKD) framework that combines a pretrained WSI image encoder and a clinical text encoder using Low-Rank Multimodal Fusion (LMF) to efficiently model cross-modal interactions during training. Each WSI is represented as a bag of patches paired with a slide-level diagnostic caption. The teacher model learns fused image-text representations for subtype classification, while the student model distills this knowledge to enable accurate image-only inference. We evaluate our method on the PatchGastric benchmark dataset and achieve at least 3.35% higher mean accuracy than state-of-the-art approaches, without relying on transformer-based fusion, multi-task learning, or large language models. The source code is available at https://github.com/helomelo1/MKD-LMF.

Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters. We present Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into a single shared cosine space without updating any text-side parameter. Our key insight is that the teacher signal requires no separate embedding model: each media sample is paired with a dense cascaded caption, and the teacher target is simply the frozen backbone's own embedding of that caption. Because teacher and student share the same backbone weights, they inhabit byte-identical geometry, and lightweight projectors plus phased LoRA adapters on the modality encoders suffice for alignment. Training combines a Matryoshka SigLIP contrastive loss with an online hybrid hard-negative miner whose negatives sharpen as the encoder improves. The recipe carries over to a 2.3B variant by swapping in a native vision-language backbone. Omni-Embed-Mini-0.9B keeps its text weights bit-identical to the backbone, so training cannot regress text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8), while extending it to five additional modalities, and is ~2.7x to 9.5x smaller than every open omni embedder we compare against. The 2.3B variant is competitive with the closed gemini-embedding-2, edging ahead of it on the overall-modality average. Models, code, data and evaluation harness are on our project page: https://omniembed.cvmbzuai.com

Large vision-language models (LVLMs) exhibit strong multimodal in-context learning (ICL) capabilities, yet this ability degrades substantially as model size decreases. Knowledge distillation offers a natural way to bridge this gap, but existing methods primarily align output distributions or hidden representations directly. Such alignment teaches the student what the teacher predicts without revealing which evidence in the complex context causally supports that prediction. Consequently, a student can imitate the teacher's answer while continuing to rely on language priors, prompt structure, or other spurious cues. To address this limitation, we introduce Multimodal Causal Distillation (MCD), a distillation framework that transfers how a strong teacher uses multimodal evidence during ICL. MCD uses structure-preserving token interventions to identify and verify causal evidence, then transfers how the teacher responds when that evidence is retained or removed. This design connects distillation to the causal patterns by which the model uses contextual evidence during multimodal ICL. Experiments across three LVLM families and seven benchmarks show that MCD improves student performance by 7.23 points on average and outperforms vanilla distillation by 4.68 points, while further analyses confirm the generalizability of these gains.

Federated knowledge must remain usable by recipients with different modalities, private architectures, and tasks. We present FedSocket, which makes recipient execution a design requirement of the exchanged model. A shared Q combines recipient-computable inputs, task-owned outputs, and ownership-aware aggregation, connecting heterogeneous private models through a common prediction interface. Private models teach local Q copies; the returned Q supports local learning and Joint inference, with only Q parameters and counts exchanged. Across six datasets, FedSocket improves missing-modality recipient accuracy over Local by 14.44 and 15.51 percentage points on MELD and UCF-51. Under matched inference capacity, Joint exceeds independent ensembles by 11.06 points in UCF-51 accuracy and 4.87 points in mean bidirectional Flickr30k R@1. Joint also improves over Q alone on all four heterogeneous endpoints, demonstrating the value of combining local and exchanged predictions. Teacher controls, sharing-path interventions, and component factorials identify the roles of supervision, sharing, and deployment. FedSocket makes exchanged knowledge directly usable from federated training to recipient inference.

Cross-modal knowledge distillation transfers knowledge from a teacher modality to a student modality. Existing feature-level alignment methods typically assume that teacher and student features reside in structurally alignable representation spaces. However, this assumption does not hold when cross-modal features are structurally heterogeneous and lack clear unit-level correspondence, such as 2D spatial visual grids and 1D temporal audio sequences, thereby limiting the applicability of feature-level alignment. To address this challenge, we propose a cross-modal distillation framework that enables effective knowledge transfer across structurally heterogeneous feature spaces via a vector-quantized codebook. Specifically, teacher features are abstracted into a set of vector-form codes regardless of their original feature structure, and the selected codes serve as concept-level anchors for student learning. Code selection is guided by both task relevance and student compatibility, allowing the student to receive transferable teacher knowledge without requiring direct unit-level feature alignment. Experimental results across diverse cross-modal distillation scenarios demonstrate the effectiveness of the proposed framework on classification and semantic segmentation tasks.

On-policy distillation (OPD) specializes pretrained video diffusion models through teacher supervision along the student's own generation trajectory. Although large video models are natural teachers, developing specialized video experts can require costly video data and training, while querying them incurs substantially higher latency than querying image experts. More readily available and cheaper to query, image experts offer a cost-effective alternative, particularly for largely temporal-agnostic capabilities such as aesthetics and OCR that admit frame-level supervision. However, heterogeneous image and video latent spaces prevent direct supervision of intermediate student states, while image experts lack cross-frame motion supervision, making temporal consistency vulnerable to frame-level improvements. In this paper, we propose MILD, a Motion-Preserving Image-to-Video Latent Distillation framework that transfers specialized image expertise while preserving pretrained video dynamics. MILD uses a learnable linear connector that aligns student latent states and predicted updates with those of image experts, enabling supervision transfer across heterogeneous latent spaces. We further constrain image-guided corrections around the pretrained student's predictions to preserve video dynamics and incorporate an optical-flow-based motion reward to improve motion quality and temporal consistency. Across specialized image experts and multiple video-student backbones, our method consistently outperforms video-teacher OPD baselines, with further studies demonstrating effective transfer across connector designs and heterogeneous architectures. These results establish image-to-video distillation as an effective route to improving video generation by drawing on the diverse and evolving capabilities of the image-generation ecosystem.

Event cameras have shown great potential for robust visual perception, yet scaling event representation learning remains challenging due to the scarcity of large-scale annotated event data. Pretrained image models provide scalable semantic supervision, but existing image-to-event methods rely on rigid pixel-wise or token-wise alignment that overlooks modality discrepancies in texture, density, and appearance, potentially causing semantic collapse and limiting transferability. To address this issue, we propose Hyper-RED, a simple, painless, and scalable image-to-event pretraining framework that transfers high-order semantic structures from images to events. Hyper-RED uses hypergraphs to model and align high-order semantic associations among multiple image and event tokens, enabling cross-modal knowledge transfer while accommodating modality-specific differences rather than enforcing rigid one-to-one correspondence. Specifically, given a paired event--image sample, Hyper-RED leverages DINOv3 to extract spatial token representations and constructs image, event, and cross-modal semantic hypergraphs, where each hyperedge connects multiple semantically correlated tokens. We further introduce a hypergraph relational distillation loss that imposes complementary intra- and cross-modal constraints, enabling the event encoder to inherit image-derived semantic organization while preserving local relational consistency and event-specific characteristics. Experiments on three tasks across five event datasets demonstrate consistent scaling from ViT-S to ViT-L and state-of-the-art performance (Fig.1). The code is available at: https://github.com/meisenwang/Hyper--RED.

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease in which early assessment remains challenging, particularly in low-resource settings where MRI is often unavailable. High-resolution ultrasound (HRUS) of the tongue offers a portable and low-cost alternative for evaluating bulbar involvement, but learning reliable diagnostic models is limited by small datasets and the difficulty of extracting robust representations from ultrasound alone. We propose AlignUS, a cross-modal knowledge distillation framework that transfers anatomical knowledge from MRI to a HRUS-based classifier while requiring only HRUS at inference time. The model combines classification loss, supervised contrastive learning, and feature-level distillation to align HRUS representations with MRI embeddings. AlignUS achieves a patient-level balanced accuracy of 0.958, macro-F1 of 0.963, and ROC-AUC of 0.990, aggregated across four patient-level cross-validation folds, with consistent improvements over HRUS baselines and cross-modal alternatives. These results demonstrate that MRI-derived supervision can substantially improve ultrasound-based ALS assessment while preserving low-cost, inference-time independence from MRI.

Unpaired cross-modal distillation transfers grade structure from histopathology into a micro-ultrasound (micro-US) encoder by aligning a pooled needle-region embedding to a frozen histopathology teacher under grade-group correspondence alone. A single objective is thereby required to serve two distinct functions: rendering patch features discriminative of tissue state, and selecting which patches enter the pooled representation. We decouple them. Weak spatial supervision derived from percentage involvement, recorded routinely at biopsy, constrains the predicted proportion of malignant tissue within each core, acting on the encoder features independently of the alignment objective. The alignment loss then operates on features that differ across a core, and attention concentrates on a subset of patches rather than remaining near-uniform. On 7,166 biopsy cores from 811 patients across seven centers under patient-level 5-fold cross-validation, the method reaches 67.1 macro AUC and 68.5 csPCa AUC, against 61.2 and 52.8 for the existing unpaired alignment method and 63.1 and 62.6 for the strongest unimodal baselines. Ablation against existing attention regularizers designed to prevent attention-uniformity collapse shows that such regularizers do not substitute for label-derived supervision: they constrain the attention distribution, whereas the signal required acts on the features that attention reads.

Parkinson's disease is a progressive neurodegenerative disorder characterised by gradual deterioration of movement control. Automated freezing-of-gait (FOG) detection supports the objective assessment of gait-related motor impairment. Two common approaches are used for FOG prediction: (i) analysing video recordings of the patient's movements and (ii) analysing data collected using inertial measurement unit (IMU) wearable sensors attached to the patient's lower limbs. Video-based approaches may suffer detection errors during continuous turning-in-place tasks because the lower limbs undergo substantial geometric self-occlusion, degrading pose-estimation accuracy. IMU-based approaches are generally less affected by visual occlusion; however, they are difficult to deploy outside clinical or laboratory settings, as the sensors must be attached securely and remain in place throughout the assessment. Motivated by this, we propose a cross-modal subspace distillation framework to mitigate the limitations of unimodal FOG detection by combining IMU accuracy with video-based practicality. We extract invariant latent topologies from a pre-trained kinematic oracle to structurally supervise a non-encoded visual architecture during training. To resolve periods of severe spatial occlusion, a dual-stream visual model probabilistically fuses skeletal graph nodes and continuous spatial pixels, dynamically shifting reliance to uninterrupted pixel boundaries as joint tracking confidence drops. Evaluated against a public, multi-modal sequence dataset of Parkinson's individuals executing continuous $360^\circ$ turns, empirical results demonstrate that applying sensory boundary topologies strictly mitigates tracking evaluation entropy. Our constrained optimisation confirms that highly precise FOG prediction bounds can be achieved over zero-wearable inference environments.

An action chunk can span several stages of a manipulation task, yet a label for its first step describes only the current stage. We introduce Chunk-Aligned Semantic Distillation (CASD), which derives semantic targets for entire action chunks. An offline vision--language model segments demonstrations into described stages. Their occupancy within each action chunk determines a weighted semantic target, including transitions between stages. A CASD generator learns to predict this target from the current observation, robot state, and task instruction. We then freeze the generator and train a policy conditioned on its predictions. The semantic branch runs once per policy query, without online VLM calls or reasoning-trace decoding. Teacher matching on annotated LIBERO training episodes is above chance for both single-stage and boundary-crossing chunks. We evaluate three Fast-WAM variants and a DreamZero integration across four benchmarks, including distribution shifts on LIBERO-Plus. Compared with published references, IDM+CASD reaches 98.9% versus 98.0% average success on LIBERO, while Uncond falls below its reference. Joint+CASD reaches 93.0% versus 90.6% on RoboTwin 2.0, and DreamZero+CASD reaches a 47.9% four-category MolmoSpaces manipulation average versus 40.7%. Performance varies across backbone integrations.

Objective assessment of Freezing of Gait (FoG) in Parkinson's disease (PD) relies predominantly on wearable Inertial Measurement Units (IMUs). While IMUs provide optimal kinematic precision, mandatory sensor attachment restricts continuous clinical deployment. Conversely, unobtrusive vision-based alternatives suffer substantial classification errors during turning-in-place tasks, where geometric self-occlusion degrades deterministic skeletal coordinates and obscures the high-frequency precursors required for FoG detection. To resolve these physical observation limits, we propose a supervised cross-modal subspace distillation framework. During optimisation, pre-trained kinematic data from IMU sensors and contextual clinical metadata act as oracles to guide a deployable visual architecture. By incorporating joint velocity and acceleration derivatives, utilising a confidence-based gating mechanism, the visual model mitigates some of the tracking errors during occlusion events. Empirical evaluations confirm this latent alignment transfers the predictive fidelity of hardware sensors directly into the visual representation, yielding $85.5\%$ accuracy, and $82.4\%$ balanced accuracy. All the while maintaining a vision only model at inference.

Multimodal Emotion Recognition (MER) systems often suffer from missing modalities in real-world scenarios. Existing methods usually generate, align, or distill missing modalities as a whole, overlooking the heterogeneous nature of the information carried by each modality. Such holistic treatment mixes inferable shared semantics with uncertain modality-specific details, yielding unstable representations and degrading robustness. To address this issue, we propose the Primitive Memory Distillation (PriMD) framework. Unlike existing methods, PriMD takes an intra-modal perspective and focuses on how different types of information within a modality differ in recoverability within each modality. PriMD first disentangles cross-modal shared semantics from modality-specific representations, and then discretizes the latter into learnable semantic primitives to construct modality-specific memory banks. When modalities are missing, PriMD is a teacher-student framework that the student model uses the shared semantics of available modalities as queries to dynamically retrieve primitives. It compensates for missing modality-specific information within a constrained memory space and aligns with the teacher model. Extensive experiments on IEMOCAP, CMU-MOSI, and CMU-MOSEI demonstrate that PriMD achieves state-of-the-art performance and consistently stronger robustness across a wide range of missing-modality settings, while mitigating the instability caused by holistic feature inference. Our code and project website are available at https://github.com/JiaqiZhang-Sengoku/PriMD and https://jiaqizhang-sengoku.github.io/PriMD/, respectively.

Depth estimation from thermal images is highly valuable for robotic applications in adverse conditions, such as nighttime and rainy weather. Recent studies have sought to transfer knowledge from RGB-based foundation models to thermal modalities, yet the rich hierarchical representations these models encode remain underutilized. To address this limitation, we propose RGB-HS, a novel framework for thermal-image depth estimation that leverages hierarchical supervision from an RGB-based foundation model. Specifically, we first replace the baseline thermal encoder with a foundational model and introduce a parallel RGB branch that also employs a foundational model as an encoder of the same architecture, taking RGB images as input. The alignment is then performed across multiple levels between the tokens of the two encoders, allowing the thermal student branch to capture both structural precision and semantic abstraction from the RGB teacher branch. Furthermore, we introduce verification to refine the alignment process by weighting tokens from the RGB branch based on RGB image quality. Extensive experiments on the popular benchmark demonstrate that RGB-HS achieves competitive performance and more effectively exploits the representational capacity of RGB-based foundation models for depth estimation on thermal images.

Multi-step language-model agents repeatedly process growing interaction histories, leading to substantial context costs. Vision--text compression reduces these costs by rendering history as images, but the resulting modality shift creates a marked capability gap. Through controlled evaluations of history recovery, matched-state decisions, and complete trajectories, we show that this gap cannot be explained by OCR quality alone. Visual-history agents exhibit systematic drift in action selection, query formulation, stopping, and evidence use, revealing an agentic policy gap. We introduce \textbf{CAPS}, a two-stage \textbf{C}ross-modal \textbf{A}gentic \textbf{P}olicy \textbf{S}elf-distillation framework that uses the same model's stronger text-history policy to supervise its visual-history counterpart. Offline trajectory self-distillation transfers successful text-policy behavior to visual-history inputs, while online policy self-distillation provides dense supervision on states visited by the visual-history policy during reinforcement learning. On SearchQA, CAPS improves over AgentOCR by 5.0% and 3.4% with 3B and 7B backbones, respectively. On full-history ALFWorld, the corresponding gains are 15.6% and 14.5%. Across settings, CAPS reduces average memory-context cost by up to 63.3% and peak cost by up to 83.4% relative to matched text-history policies. These results show that explicit cross-modal policy self-distillation can preserve agent capability under vision--text compression. Our code will be made publicly available in a future release.

World Action Models (WAMs) improve robot policy learning by jointly modeling future visual dynamics and actions. However, their scalability and generalization remain constrained by their reliance on costly expert demonstrations. We challenge this by asking whether future supervision for WAMs must originate from target-task expert trajectories. In this paper, we propose Vid2WAM, an offline distillation framework that transfers visual diffusion priors from a large video foundation model into a compact WAM student. Given an observation and language instruction, Vid2WAM distills supervision through two complementary channels: task-conditioned future rollouts directly supervise the student's future prediction branch, while an inverse dynamics model recovers embodiment-specific pseudo-actions for action learning. To robustly integrate synthetic and real supervision, we introduce source-aware residual action adaptation that learns source-specific corrections around a shared action backbone and mitigates interference from noisy pseudo-actions. During inference, both the video teacher and inverse dynamics model are discarded, leaving only the WAM student for efficient deployment. Simulation and real-world experiments demonstrate that Vid2WAM improves novel-task generalization and data efficiency under limited expert demonstrations while preserving low-latency inference.

To address the limitations of video-based emotion recognition under ambiguous or socially masked behavioral cues, as well as the poor deployability of physiological signals, this paper proposes a reliability-aware physiology-to-video knowledge distillation framework, termed BioKD. The proposed framework leverages physiological signals as privileged information during training to guide a video-based student model in learning deep affective representations, while relying solely on non-intrusive video inputs at inference time. To cope with the high noise and instability of physiological teacher supervision caused by inter-subject variability, signal artifacts, and temporal inconsistency, BioKD incorporates a sample-wise reliability-aware gating mechanism together with a progressive distillation strategy. By adaptively regulating the strength of knowledge transfer, the framework suppresses negative transfer induced by unreliable physiological supervision and enables more stable cross-modal distillation. Experiments on DEAP and AMIGOS show that BioKD consistently outperforms representative baselines under both trial-wise and subject-wise evaluation protocols for valence and arousal recognition. For example, BioKD achieves 68.01% on DEAP (trial-wise arousal) and 65.29% under the more challenging subject-wise setting, demonstrating improved performance under a subject-independent evaluation setting. Further analyses show that BioKD effectively mitigates overconfident teacher errors and outperforms an entropy-only weighting strategy, confirming the importance of explicitly modeling supervision reliability. In addition, BioKD introduces no additional inference-time overhead relative to the same video student architecture and removes the need for physiological sensing and multimodal synchronization.

Optical microscopy of particle and fiber dispersions involves interpreting subtle visual cues influenced by specimen morphology, chemical composition, magnification, and illumination conditions. We introduce an artificial intelligence (AI) distillation framework that extracts semantically rich image embeddings from microscopy images using semantic anchors. A multimodal teacher combines each image's visual embedding with three text embeddings representing illumination modality, magnification, and specimen identity and morphology. Generated by LongCLIP's extended-context text encoder, this yields a 2304-dimensional block-structured teacher vector whose component blocks remain physically interpretable throughout training and inference. A student vision transformer (ViT) with a multi-layer perceptron (MLP) decoder is trained to reconstruct this teacher vector from the image alone, minimizing a mean absolute error (L1) loss that enforces coordinate-level fidelity to the teacher's block structure. A cross-entropy term over pseudo-classes derived from HDBSCAN clustering of the teacher embedding space acts as a collapse-prevention regularizer, enforcing inter-cluster separation without requiring contrastive negative mining. At inference, the student operates on image input alone, producing compact embeddings that recover the full semantic content of the teacher vector. The framework achieves approximately 80% pseudo-class validation accuracy and 75% Recall@1 on fine-grained specimen description labels under leave-one-out nearest-neighbor retrieval. These results demonstrate that semantic anchoring enables a vision-only student to acquire richer and more interpretable representations than image-only training, with direct applicability to retrieval, classification, and exploratory analysis of heterogeneous particle and fiber dispersions.

The deployment of Vision-Language Models (VLMs) in critical domains like disaster management requires high-quality multimodal datasets, especially for transferring knowledge via Data-Free Knowledge Distillation (DFKD). However, existing datasets in this domain either entirely lack descriptive text, such as Incidents1M, or suffer from severe text-image semantic misalignment, such as CrisisMMD. In this work, we present a novel methodology to construct and automatically validate a large-scale multimodal dataset for disaster response. Starting from the vision-only Incidents1M, we successfully recovered 100,000 images and generated high-fidelity textual descriptions using two distinct Qwen3.5 architectures: a 4B dense model and a 35B Mixture-of-Experts (MoE) model. To ensure the generated captions provide reliable semantic anchoring for DFKD, we introduce an image-blind LLM-as-a-Judge validation pipeline leveraging Qwen3.5-9B. By intentionally obscuring the original image from the judge, this evaluator accurately simulates the modality gap of the student model during data-free distillation. Our evaluation across 173,179 label pairs demonstrates a high semantic agreement (78.65/100) between the two architectures. Furthermore, the automated evaluation reveals a conservative captioning behaviour, characterized by a high Precision (77.6%) and low Recall (46.0%). This minimizes the false positive noise, while simultaneously exposing underlying human annotation inconsistencies in the original ground truth. This work provides a scalable, LLM-validated multimodal dataset and a reproducible framework to advance cross-modal knowledge distillation.

Cross-modal knowledge distillation can transfer diagnostic knowledge from a strong but costly teacher modality to a cheaper and more deployable student modality. In medical image analysis, however, the two modalities are often unpaired: they are collected from different patient cohorts and occupy geometrically incompatible feature spaces. This makes instance-level distillation invalid and direct feature matching unreliable. To address these challenges, we propose Shared Semantic Codebook Distillation (SSCD), which compares teacher and student representations through a shared discrete codebook. Each image is represented as a distribution over a common, modality-agnostic vocabulary, and knowledge is transferred by aligning these distributions across modalities, both globally and class-conditionally, without requiring paired samples or directly comparable raw features. The codebook is evolved online by exponential moving average and kept diverse through entropy regularization and dead-code restart. At inference, all teacher-side and codebook modules are discarded, leaving only the student encoder and classifier. On two heterogeneous unpaired settings, OCT-to-fundus retinal disease classification and CT-to-chest-X-ray pneumonia classification, SSCD improves the student from 64.5 to 70.2 macro-F1 and from 73.8 to 76.3 macro-F1, respectively, outperforming all evaluated distillation baselines on both settings. Code and pretrained models are available at https://github.com/DillanImans/SSCD-unpaired-distillation

Unlike large language models (LLMs) that exhibit strong reasoning capabilities, vision-language models (VLMs) struggle with visual reasoning, even on geometry problems that admit equivalent text, diagram, and combined diagram+text views. We show that these views often elicit different behaviors: a model may solve a problem from text but fail on the corresponding diagram, or succeed visually while failing textually. This inconsistency suggests that different views expose complementary reasoning paths and failure modes that standard multimodal post-training does not fully exploit. To study and exploit this phenomenon, we construct ODA-Data, a high-quality paired multimodal geometry dataset with text-dominant, image-dominant, and combined image+text views of the same problems, together with splits for training and evaluating modality-dependent reasoning behaviors. We then develop Modality-Informed Reciprocal Reasoning Optimization (MIRROR), a reinforcement learning approach for improving multimodal reasoning via self supervision. For each problem, MIRROR evaluates the model under all views, selects the best-performing view as a teacher, and trains other views with a reverse-KL objective towards the teacher. Across reasoning benchmarks that evaluate on geometry problems, MIRROR improves over standard RL and yields more accurate and consistent behavior across modalities

While large audio-language models have achieved remarkable progress in auditory perception, they still lag behind text-based large language models in deep logical reasoning, primarily due to the scarcity of high-quality audio reasoning data. To bridge this gap, we propose X$^3$-OPD, a cross-modal on-policy distillation framework that transfers reasoning capabilities from a powerful text teacher to an audio-language student. During training, the student generates reasoning trajectories conditioned on its own acoustic perception, while the teacher provides token-level guidance using matched textual inputs and verified answers. We further construct a three-tier symmetric corpus covering textual reasoning rendered into speech, audio-event reasoning grounded in complex acoustic scenes, and spoken-dialogue reasoning involving paralinguistic cues. This design extends cross-modal distillation beyond textually recoverable content to reasoning grounded in non-linguistic events, prosody, and conversational context. Experiments on MMSU, MMAU, BIG Bench Audio, and MMAR demonstrate that X$^3$-OPD substantially improves audio-grounded reasoning and chain-of-thought quality while largely preserving the model's existing capabilities under domain shift.

Multimodal based approaches often outperform single modality approaches in downstream tasks as the different modalities provide complementary information, yet acquiring paired clinical data remains a significant challenge in real world scenarios. While cross-modal knowledge distillation addresses this, existing methods often struggle with large modality gaps and the propagation of noise from uncertain source-domain predictions. To overcome these challenges, we propose UnDA, an anchor-guided framework for unpaired cross-modal distillation. Our approach introduces a backbone-agnostic Alignment Module that extracts semantically structured class tokens via an attention based pooling mechanism. To ensure robust knowledge transfer, we propose Uncertainty-Weighted Optimal Transport (UCT-OT), which dynamically weights feature-level alignment based on prediction confidence, effectively suppressing noisy supervision. Furthermore, a per-class ProtoNCE objective maintains stable prototype memories to enforce global discriminability across unpaired batches. Evaluations on representative segmentation tasks under strictly unpaired settings show consistent improvements in accuracy and boundary precision in the target modality, demonstrating that meaningful structural knowledge can be transferred across heterogeneous data sources without paired datasets.

Omni-modal models provide a unified interface for text, images, and audio. However, improving these abilities together remains difficult, as post-training on pooled multimodal data often fails to preserve the strengths of modality teachers. On-policy distillation (OPD) has recently become popular in model post-training. It samples responses from the current student and compares the teacher's and student's next-token distributions along those responses, yielding dense supervision while reducing the mismatch between training and inference. Despite these advantages, standard OPD does not readily extend to several modality teachers. Their guidance may favor conflicting changes to the shared model, while matching each teacher's next-token distribution can prevent the student from moving beyond that teacher. To address these challenges, we propose On-Policy Omni Distillation (OPOD), which consolidates text, image, and audio teachers into one omni model. OPOD routes each response to the corresponding teacher, controls the teachers independently, and applies guidance only when the teacher assigns a higher probability to the generated token. The selected teacher also evaluates answer confidence and whether the reasoning increases support for the answer. Extensive experiments on twelve benchmarks show that OPOD achieves the best average at three model scales, reaching 70.8, 51.7, and 46.2 and outperforming the strongest comparator by 2.1, 1.8, and 1.7 points. At 30B, it surpasses the base model and pooled RL training on all twelve benchmarks, and ranks first or second on eleven even when the teachers are included. Only the student is retained for deployment.

Visual grounding with multimodal large language models is commonly formulated as autoregressive coordinate generation, where a model outputs bounding-box coordinates as text given an image and a referring-expression prompt. While this interface is simple and compatible with instruction following, it introduces a mismatch between training and evaluation: training optimizes token-level likelihood over coordinate strings, whereas grounding quality is measured by geometric overlap. We propose IoU-PD, an IoU-aware privileged distillation method for coordinate-generating multimodal large language models. IoU-PD uses ground-truth boxes not only as coordinate targets, but also as privileged training-time guidance. During training, the student receives the original image and prompt, while a frozen teacher receives a box-marked image and an augmented prompt that indicates the marked region. The student is trained with a supervised fine-tuning anchor and a privileged distillation loss whose token weights reflect both geometric importance and teacher reliability. At inference time, IoU-PD requires no box overlay, privileged hint, teacher branch, or additional prediction module. Experiments on standard referring-expression grounding benchmarks show consistent region-level improvements over strong coordinate-generating baselines, demonstrating that ground-truth boxes can provide useful privileged guidance beyond serving as coordinate labels.

Cross-modal distillation from Vision Foundation Models (VFMs) to LiDAR backbones has recently emerged as a self-supervised pretraining strategy that reduces reliance on dense point-wise annotation for 3D scene understanding. However, existing distillation pipelines typically treat the VFM as a frozen feature source and train a heterogeneous 3D backbone to match fixed image embeddings, forcing the student to bridge both the modality gap and the cross-architecture gap between dense ViT token representations and sparse 3D encoders. We propose TOLiD, a self-supervised pretraining method for LiDAR representation learning that addresses this gap by coupling a LiDAR backbone with a student Vision Transformer (ViT) initialized from a frozen VFM teacher and applying supervision over compatible patch-token representations. TOLiD converts the set of point features within each image patch frustum into a token using Frustum Pooling followed by Frustum Attention, and performs token-level distillation with visibility masking. For LiDAR-only deployment, we lift token features back to per-point representations using masked bilinear sampling to avoid patches that have limited LiDAR points. We extensively evaluate TOLiD on five heterogeneous LiDAR datasets and four cross-sensor adaptation pairs, demonstrating improved transfer with frozen backbones and lightweight heads.

Building pixel-level correspondence between event and image data is a fundamental task for multi-sensor systems. However, existing cross-modal matching methods are largely restricted by their reliance on either matching labels or strictly aligned hardware, which limits them to unlabeled and unconstrained real-world scenarios where neither matching ground truth nor prior sensor relationships are available. To address this, we propose a novel two-stage training paradigm. First, we leverage large-scale data to perform label-agnostic distillation pretraining, upgrading optimization objectives with distribution-based and contrastive losses to learn highly generalizable representations. Second, to tackle unlabeled and unconstrained downstream data, we introduce an epipolar-guided self-distillation framework. By utilizing consistency verification to isolate robust matches and incorporating geometric confidence derived from an external epipolar prior, our model can effectively self-evolve directly on target domains without any supervision. Furthermore, we introduce a rigorous cross-modal evaluation benchmark based on TUM-VIE, featuring physically separated cameras with distinct intrinsic parameters and resolutions. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance on both MVSEC and TUM-VIE pose estimation tasks. The source code and benchmark will be made publicly available at https://github.com/ZhonghuaYi/nexus2-official.

Accurately estimating urban carbon emissions is critical for sustainable urban planning, yet many existing approaches remain difficult to apply consistently across cities due to data-source heterogeneity and the lack of fine-grained semantic-temporal context in remote sensing data. We propose CarbonCLIP, a task-oriented multimodal distillation framework that improves satellite-based carbon emission prediction by transferring contextual knowledge into a unified satellite representation through dual-branch contrastive learning. Unlike conventional methods that rely on static visual features, CarbonCLIP explicitly bridges the gap between top-down satellite views and ground-level human activities. Specifically, the spatial branch uses fine-grained textual descriptions automatically generated from street-view images by Large Multimodal Models (LMMs) to provide semantic priors reflecting building functions, infrastructure, and urban activities, while the temporal branch employs a month encoder to encode temporal priors associated with monthly emission variation. CarbonCLIP requires multimodal data only during the pretraining phase; during inference, it relies solely on satellite imagery, thereby supporting scalable deployment when ground-level data are unavailable at inference. Experiments on Beijing and Singapore demonstrate that CarbonCLIP outperforms baselines in both study cities. The results validate that our method effectively transfers multimodal knowledge into satellite representations, offering a robust solution for satellite-based urban carbon modeling.