Cross-Modal Knowledge Distillation

Plasma disruption is a critical threat to tokamak safety. Existing data-driven predictors mainly rely on time-series diagnostic signals, while visible images provide complementary spatial cues including plasma deformation, local brightening, and radiation-structure evolution. Although the image modality improves the model's discriminative capability, it also substantially increases the computational cost during inference. To address this issue, we propose a hierarchical multi-to-single-modal knowledge distillation framework for disruption prediction on a synchronized EAST multimodal dataset. During training, visible images and time-series signals are used to train a multimodal teacher, which learns disruption precursor representations through Transformer-based encoders and a prototype-guided spatiotemporal hypergraph module. During inference, only the time-series student is retained, with multimodal knowledge transferred through graph-structure-level, representation-level, and decision-level distillation. On the 640-discharge EAST dataset, the results demonstrate that the proposed framework can preserve the discriminative advantages of multimodal learning while substantially reducing inference cost, and providing an effective route for efficient disruption prediction in EAST. The source code of this paper will be released on https://github.com/Event-AHU/OpenFusion.

While knowledge distillation (KD) is widely adopted for training lightweight models by leveraging supervision from larger teacher models, relying solely on output token distributions has proven insufficient for compressing Multimodal Large Language Models (MLLMs). Since output tokens are a byproduct of the model attending to visual inputs, prior works have explored explicitly distilling attention to provide a direct supervisory signal. While promising, the precise utility of which attention signals to distill remains under-explored. In this work, we challenge the conventional reliance on prompt-to-vision attention by revealing that downstream performance correlates strongly with response-to-vision attention similarity to the teacher, but negligibly with that of prompt-conditioned attention. Furthermore, we observe that attention distributions exhibit significant variance across individual tokens, indicating that a uniform distillation objective is suboptimal. To this end, we introduce Token-level Response-visual Attention Guidance (TRAG), a distillation objective that 1) shifts the focus to response-to-vision signals and 2) employs token-specific objectives by adaptively weighting the Kullback-Leibler divergence based on attention entropy, effectively guiding the student to mirror the teacher's precise visual focus. Extensive experimental results on multiple benchmarks demonstrate that TRAG significantly outperforms prior distillation baselines.

On-policy distillation (OPD) has recently emerged as an effective post-training paradigm by providing supervision on student-generated trajectories. However, existing OPD methods for multimodal reasoning usually rely on a static teacher routing, assigning each sample to a single teacher based on modality or task type. This ignores that visual grounding and abstract reasoning may dominate different decoding steps, making a single teacher insufficient for the full trajectory. To this end, H-OPD is proposed as a confidence-aware heterogeneous multi-teacher OPD framework for multimodal reasoning. By verifying the complementarity of heterogeneous teachers in the same reasoning process, H-OPD replaces task or sample level teacher routing with token-level teacher arbitration along the shared student trajectory. H-OPD employs vision-to-language description transfer to enable text-only teachers to access key visual semantics, and uses a confidence-aware arbitration mechanism to dynamically combine vision-language teacher and text-only teachers at each token. Extensive evaluations over 11 widely-used reasoning benchmarks showcase the superior performance of our method.

Automatic understanding of dynamic 4D point clouds, the 3D-point sequences captured over time by depth sensors and LiDAR, is central to robotics and embodied perception. Yet annotating them densely is expensive, making self-supervised pretraining the natural route to transferable representations. Existing pretext tasks, however, are almost entirely intra-modal, and the few methods that transfer knowledge from 2D foundation models rely on a single global embedding per clip, discarding the rich per-patch semantics that these models compute. To address this gap, we propose Cross4D-JEPA, a teacher-student method that distills a frozen 2D foundation model, an image model DINOv2, or a video model V-JEPA 2, into a 4D point encoder. The proposed method combines (1) a dense cross-modal correspondence that maps every 3D point to the teacher patch feature it projects to, and (2) a per-point objective that trains the student to match these features in latent space with no masking, negatives, or decoder. We evaluate Cross4D-JEPA on four benchmarks, MSR-Action3D, DeformingThings4D, NTU-RGB+D 60, and HOI4D, against intra-modal and global cross-modal baselines. Experimental results show that, under a matched protocol, the proposed method consistently outperforms intra-modal and global cross-modal baselines across the four benchmarks and is competitive with heavier published 4D methods; further analysis attributes this gain primarily to the granularity of the correspondence rather than the teacher modality. Beyond recognition accuracy, the dense representation learned by Cross4D-JEPA transfers across domains, improves label efficiency, and improves full-label fine-tuning under the same training budget, while a 13x smaller encoder matches a heavyweight pooling backbone.

Surface electromyography (sEMG)-based gesture recognition has emerged as a promising technology for natural human-computer interaction. However, its practical deployment remains challenging due to severe performance degradation caused by feature distribution discrepancies across different subjects and recording sessions. Although domain adaptation (DA) techniques are commonly employed to mitigate such discrepancies, conventional methods often struggle to effectively aligning sEMG features, primarily due to their inherent stochasticity and the scarcity of labeled data. To address these limitations, this paper proposes a novel Pressure-Guided Unsupervised Domain Adaptation (PGUDA) framework, which leverages the robustness and stability of pressure signals to introduce a cross-modal knowledge distillation strategy that transfers consistent physical semantics across modalities. Specifically, a teacher network trained on pressure signals guides an sEMG student network on unlabeled target domains, thereby regularizing the representation learning process with transferable and modality-invariant knowledge. Extensive experiments conducted on a self-collected multimodal dataset involving eleven subjects validate the effectiveness of the proposed PGUDA framework. The results demonstrate that our proposed PGUDA achieves leading performance in both cross-subject and cross-session classification tasks, achieving average accuracies of 58.08% and substantially outperforming existing DA approaches. Notably, PGUDA exhibits remarkable label efficiency: it attains classification accuracy comparable to fully supervised benchmarks while requiring only 5% of labeled data for teacher network training. This framework offers a robust and data-efficient solution that can significantly reduce the calibration burden in practical sEMG-based gesture recognition systems.

Color Fundus Photography (CFP) offers a low-cost and non-invasive route for ischemic heart disease (IHD) screening, but current studies are limited by scarce public benchmarks and ineffective fusion of retinal images with sparse clinical variables. We propose IDNet, a multimodal framework with a Cross-Modal Distillation Aggregator (CDA) that uses learnable queries to sequentially integrate left-eye, right-eye, and clinical features, mitigating the imbalance between high-dimensional visual features and low-dimensional tabular inputs. We also construct a reproducible UK Biobank benchmark with open-source curation and quality-control pipelines, yielding 50,410 images from 25,205 subjects. On this benchmark, IDNet outperforms image-only, clinical-only, and several multimodal baselines, and CDA consistently improves multiple visual encoders as a plug-in fusion module.

Video-based remote physiological measurement (RPM) is highly accessible but remains fragile under varying illumination, skin tones, and motion. Radio frequency (RF) radar is largely invariant to illumination and appearance, providing complementary cardio-respiratory micro-motion cues; however, requiring radar at inference is often impractical due to its limited ubiquity and deployment overhead. We propose RPM-Distill, a physiology-guided cross-modal distillation framework that leverages synchronized radar only during training while retaining video-only inference. Our key observation is that although RGB and RF waveforms differ in sensing physics and time-domain morphology, they share similar latent periodic rhythm in the frequency domain. We thus distill physiology-structured spectral evidence to improve robustness, via losses that (i) anchor the fundamental peak, (ii) match the off-peak background distribution, and (iii) preserve spectral morphology and sharpness. To avoid negative transfer under sample-level teacher quality and alignment uncertainty, a spectral policy network predicts sample-level distillation gates and component weights from the student--teacher spectral relation map, learned with a meta bilevel objective on a small labeled validation split. Through extensive experiments in challenging conditions and cross-dataset settings, RPM-Distill brings 81% MAE and 21% correlation improvement over unimodal baselines. Code is at https://github.com/WJULYW/RPM-Distill.

Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at https://github.com/lliangthomas/lavid.

Multi-modal fusion and multi-model ensembling are prevalent in enhancing the performance of 3D semantic segmentation. Despite the impressive performance, these methods either rely on auxiliary input signals or suffer from costly computational expense. To efficaciously enhance the segmentation performance without introducing intolerable costs, we propose to transfer the rich knowledge from the multi-modal model (i.e., point clouds and images) and multiple model experts to the point-cloudbased network through knowledge distillation. Specifically, we present Information-oriented Heterogeneous Distillation (IHD) to help the uni-modal model absorb the complementary knowledge from the multi-modal teacher. We design the Information-Oriented Filtering (IOF) strategy to select informative images from the continuous image sequence for multi-modal fusion. This practice can boost the performance of the multi-modal teacher, thus benefiting the learning of the student. Besides, as opposed to vanilla model ensembling that requires the separate training of each expert, we propose Adept Snapshot Distillation (ASD). ASD treats the freely available model snapshots generated during the training phase as multiple experts, which significantly reduces the training cost for model ensembling. For each expert teacher, it only provides supervision to the student in the class where it is adept. The resulting Heterogeneous and Adept Snapshot Knowledge Distillation, dubbed HAS-KD, attains state-of-the-art results on ScanNetV2 and S3DIS datasets. HAS-KD can be seamlessly integrated into contemporary 3D segmentation algorithms and bring considerable gains without introducing extra inference burdens. The code will be made publicly available upon publication.

Event-based object Detection (EvDet), as a biologically inspired visual perception paradigm, demonstrates superior performance in scenarios demanding high temporal resolution and a wide dynamic range. Nevertheless, the inherent sparse representations and inadequate visual semantics of event data result in a considerable performance disparity between EvDet and frame-based object detection. Previous works attempt to alleviate this cross-modal discrepancy through knowledge distillation, yet they only focus on spatial visual semantics or pair-wise relational information, thus limiting performance in more complex scenarios. To address this challenge, this paper proposes M^2C-EvDet, a Multi-domain and Multi-order Cross-modal knowledge distillation framework for EvDet. Built upon frequency learning and hypergraph computation, M^2C-EvDet integrates two specialized modules: Adaptive Frequency-Decoupled Feature Distillation (AF^2D^2) and Multi-Order Relational Distillation (MORD).

Egocentric video understanding is inherently limited by the narrow perspective of wearable cameras: a single viewpoint, a single modality, a single model cannot capture the full richness of human action. We argue that a truly expressive egocentric representation must subsume complementary knowledge across viewpoints, modalities, and foundation model representations, yet remain deployable from egocentric video alone. To this end, we introduce a hierarchical multi-teacher distillation framework that produces UNIEGO, a unified egocentric encoder trained with nine teachers spanning ego-exo viewpoints, RGB, depth, and skeleton modalities, and four foundation models. Rather than distilling directly from heterogeneous teachers whose incompatible architectures and feature geometries induce conflicting gradients, our framework interposes a layer of representation-specific Proxy models that translate diverse teacher knowledge into a homogeneous egocentric space. A second distillation stage, Selective Proxy Distillation (SPD), then adaptively selects, for each training sample, the subset of proxies that are both correct and confident, distilling exclusively from reliable supervision and suppressing erroneous signals. SPD is further stabilized by initializing UNIEGO as a learned convex combination of proxy parameters, placing the unified model in a well-conditioned region of the loss landscape before distillation begins. UNIEGO achieves state-of-the-art performance across three egocentric video understanding tasks - action recognition, video retrieval, and action segmentation on three challenging ego-exo benchmarks, outperforming naive multi-teacher distillation baselines and demonstrating that structured, proxy-mediated knowledge transfer yields richer and more discriminative egocentric representations.

Leveraging Vision Foundation Models (VFMs) for camera-to-LiDAR knowledge distillation offers a promising solution to the scarcity of annotated data needed to represent the immense geometric and kinematic diversity of real-world autonomous driving (AD). However, current approaches typically treat VFMs as black-box teachers, relying exclusively on frame-wise feature similarity. Consequently, they do not fully exploit the teacher's layer-wise semantic structure and global context, as well as the rich spatiotemporal information inherent in LiDAR sequences. We propose HilDA, a self-supervised pretraining framework for LiDAR backbones that better captures the semantic what and geometric where needed for driving tasks. HilDA combines hierarchical distillation comprising multi-layer distillation for progressive semantic alignment and global context distillation for scene-level semantics, with a temporal occupancy diffusion objective promoting spatiotemporal consistency. Models pre-trained with HilDA achieve state-of-the-art results on cross-modal distillation benchmarks and outperform models trained via prior distillation approaches on 3D object detection, scene flow, and semantic occupancy prediction. Code available at: https://maxiuw.github.io/hilda.

Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly an order-of-magnitude inference cost from multi-step diffusion. We find this cost yields limited direct benefit. On ThinkMorph, removing or noising VTs barely changes accuracy across nine benchmarks. Once rendered, attention concentrates on the VT regardless of content. Yet a KL diagnostic shows that conditioning on a privileged VT trace shifts the model's completion distribution. This suggests the generation pathway encodes useful reasoning beyond the rendered pixels. Motivated by this gap, we propose Visual On-Policy Self-Distillation(Visual-OPSD). Teacher and student share identical weights but differ in context: the teacher sees privileged VTs while the student sees only the question. Token-level JSD distillation on on-policy student trajectories transfers the teacher's reasoning to a text-only student. Across nine benchmarks, Visual-OPSD improves over its generative teacher by $+3.40$pp with $14.3\times$ speedup (10.0s vs. 142.8s per sample) and outperforms same-scale VLMs by $+63.83$pp on VSP. A Gaussian-noise control ($+0.40$pp vs. $+10.28$pp for real VTs) and $58.4\%$ closure of the KL gap confirm that gains come from the semantic content of the generation pathway.

Scalable, all-weather place recognition increasingly relies on heterogeneous radar place recognition to bridge diverse hardware platforms. A notable application is matching queries from cost-effective 4D automotive radars against high-fidelity reference maps built by dense spinning radars. This process is fundamentally limited by the extreme sparsity (and narrow field-of-view) of the 4D sensor, which captures only a fraction of the structural density present in the spinning radar database. Prior efforts address this issue by unifying different radar signals. That is, projecting both signals into a common representational space. Yet, they suffer performance degradation in multi-session environments. In this paper, we propose spatially-stratified distillation (SSD); a strategy that replaces standard uniform distillation with an asymmetric spatial alignment derived directly from physical radar returns. In regions where both radars exhibit overlapping returns, SSD enforces strong feature alignment. Crucially, in sparse regions where the 4D student lacks returns but the teacher contains valid structure within the shared field of view, SSD applies heavily discounted distillation weights. Extensive evaluations of the recent HeRCULES dataset demonstrate that SSD significantly outperforms prior place recognition methods, achieving state-of-the-art results on its challenging dynamic sequences.

Spiking neural networks (SNNs) are brain-inspired, event-driven models that compute with sparse spikes, which enables highly efficient visual perception in resource-constrained embodied AI models. The emergence of Spiking-Transformer models with spike self-attention has substantially improved the learning capacity of pure SNNs. Although SNNs are energy efficient, their performance is still limited by the spike-based architecture and optimization challenges, as standard gradient descent rules cannot be directly applied. Recently, vision-language models (VLMs) have shown rich multi-modal knowledge representation capabilities for visual perception. Thus, it is promising to leverage VLMs for better Spikformer training. To this end, we present VL2Spike, a novel spike-based knowledge distillation (KD) framework that bridges multi-modal knowledge from VLMs with compact Spikformer models. This design enhances the learning capacity of Spikformer models while preserving their energy-efficiency merits, thereby offering a practical pathway toward low-power robotic perception. Our VL2Spike brings two key technical contributions. To align with spiking dynamics, we first propose spatial-temporal visual spike (SVS) distillation, which achieves (1) shared manifold alignment between VLM image features and spike tokens, and (2) warm-started temporal consistency on membrane potentials and spike rates. We then design a novel spike prototype-guided linguistic (SPL) distillation strategy that aligns Spikformer's class prototypes and logits with promptable VLM text embeddings. Extensive experiments show that VL2Spike achieves 6.81% gain across three static datasets with only 15.7% energy consumption. It also exhibits strong generalization capacity on robotic visual place recognition (VPR) with a gain of 6.63%, highlighting its potential for low-power perception in embodied AI.

Semantic 3D occupancy provides a voxelized world state for autonomous driving and robot decision making, but object and rare-class errors can affect free-space interpretation, collision checking, and temporal state propagation. We show that a common VLM strategy, aligning 3D voxel or object features with crop-caption embeddings, improves text-space similarity without reliably improving closed-set occupancy mIoU. Motivated by this mismatch, we propose VISA, a training-time semantic auditing approach for existing occupancy world models. VISA queries an offline VLM on a representative crop of each physical object instance, obtains a structured audit with class hypotheses, plausible confusions, reliability, attributes, and evidence, and propagates it along the object track. The audit is grounded to matched 3D object voxels and distilled into semantic logits through reliability-weighted taxonomy, attribute-factor, and scene-level audit graph losses, while inference remains unchanged and requires no VLM. On nuScenes, averaged across three runs, VISA improves OccWorld from 19.06 to 20.05 mIoU and GaussianWorld from 21.36 to 21.91 mIoU; on GaussianWorld, object mIoU improves from 18.18 to 19.16 and rare-class mIoU from 15.60 to 16.79. These results suggest that VLMs are better suited to closed-set occupancy as reliability-aware semantic auditors than as generic caption-embedding targets.

Gait-based Parkinson's disease assessment increasingly relies on heterogeneous sensors, but clinical systems rarely collect all modalities simultaneously. New sensors may arrive through device upgrades, protocol changes, or multi-center deployment, while historical patient data are often unavailable because of privacy and storage constraints. This modality-incremental setting faces three challenges: unreliable cross-modal distillation, modality-specific statistical shifts, and reduced plasticity after preservation. We propose MOSAIC, a compact continual learning framework. First, we identify the Toxic Teacher phenomenon and introduce Modality-Specific Warm-Up to stabilize newly learned modality representations before distillation. Second, we propose a statistics-decoupled MSBN architecture that isolates sensor statistics while maintaining a shared semantic backbone. Third, we design a curriculum-guided repulsive objective for Plasticity Recovery, preserving legacy knowledge while recovering modality-specific capacity. Experiments on three multimodal Parkinson's gait datasets show that MOSAIC improves final performance and mitigates forgetting. Project code is available at: https://github.com/minlinzeng/MOSAIC_Modality-Specific-Adaptation-for-Incremental-Continual-Learning-in-PD-Gait-Assessment.git

Transfer learning from large-scale RGB foundation models to infrared (IR) imagery through knowledge distillation (KD) remains challenging due to fundamental differences in image formation physics. We investigate the spectral structure of the RGB--IR modality gap and observe that feature divergence is not uniform across spatial frequencies: low-frequency components (shape, layout) show greater cross-modal alignment than high-frequency components (texture, fine edges), which reflect modality-specific characteristics. Based on this analysis, we propose FreqKD, a frequency-decoupled distillation framework that applies asymmetric supervision adapted to each band's cross-modal consistency. The method employs strict mean squared error (MSE) on the low-frequency band to preserve shared structural information and a relaxed log-MSE loss (weighted at 0.1) on the high-frequency band to provide edge guidance while tolerating texture differences. Spectral divergence analysis on 500 paired samples shows that high-frequency divergence exceeds low-frequency divergence by a factor of 2.4x on average across all analysed transformer layers. On KAIST multispectral pedestrian detection, FreqKD achieves 64.1 mAP50, improving 2.4 points over the DINOv2 baseline. The learned representation transfers across datasets (FLIR ADAS, +2.1 mAP50), tasks (MFNet segmentation, +1.85 mean intersection-over-union), and architectures (ResNet-50, +1.0 mAP50). Code is available at: https://anonymous.4open.science/r/freq_decoupled_kd-5E5A

Cross-modal knowledge distillation (CMKD) studies how a (large) teacher model trained on one type of data (e.g., images) can guide a (smaller) student model building on another type of data (e.g., text/audio). Existing CMKD methods often require paired multi-modal data with aligned semantics, but obtaining such paired data are often costly and impractical. To mitigate this limitation, we develop a new CMKD framework for the more challenging setting where paired data are unavailable. In particular, we establish a cross-modal distributional relationship between teacher and student models, which reveals two fundamental quantities governing effective distillation: feature alignment and label alignment. These quantities characterize semantic discrepancy between modalities at the levels of representation and prediction distributions, respectively. Motivated by this insight, we propose a principled framework, with theoretical guarantees, that enables effective cross-modal knowledge distillation by aligning distributions rather than individual samples. Extensive experiments across a wide range of multimodal benchmarks show that our framework is highly effective in both unpaired and paired data settings, improving significantly over prior work.

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

While vision-language models excel at general multimodal understanding, they still struggle with visual spatial planning. We attribute this to a perception-reasoning modality gap: visual planning requires models to infer latent state structures from pixels and then reason over the recovered structure to produce valid actions, whereas symbolic planning directly leverages explicit objects and constraints. This creates dual bottlenecks in visual state recovery and multi-step planning. To address this, we propose MGSD, a two-stage modality-gap-aware self-distillation framework. First, a cold-start grounding stage equips the visual student with reliable state representations, minimizing early perception noise. Second, a privileged teacher transfers planning capabilities via on-policy distillation, using explicit symbolic states to supervise the student's own visual rollout prefixes. Crucially, symbolic data is used strictly during training, leaving inference purely visual. Experiments on visual planning benchmarks show that MGSD consistently improves visual planning across both 4B and 8B backbones, raising the macro average by 19.3% and 18.4%, respectively. The resulting models narrow the gap to symbolic-input upper bounds, while ablations and diagnostics confirm that the improvement comes from both visual state recovery and optimal-path reasoning. These results suggest that modality-gap-aware self-distillation improves not only how models perceive actionable states, but also how they plan over the inferred structure. Code is available at https://github.com/Oranger-l/MGSD.

Clinical diagnosis often requires combining imaging with physiological measurements, yet deployed models typically operate on unimodal data. We present PaCX-MAE, a cross-modal distillation framework that injects physiological priors into chest X-ray (CXR) encoders while remaining strictly unimodal at inference. PaCX-MAE augments in-domain masked autoencoding with a dual contrastive-predictive objective, aligning CXR representations with paired ECG and laboratory embeddings. Extensive evaluation across nine benchmarks demonstrates consistent improvements over domain-specific MAE, particularly on physiology-dependent tasks (e.g., +2.7 AUROC on MedMod; +6.5 F1 on VinDr). The method proves highly label-efficient in the 1% regime and preserves anatomical fidelity, achieving parity with MAE on segmentation tasks. Zero-shot and attention analyses confirm that PaCX-MAE successfully learns to attend to physiological indicators, such as the cardiac silhouette, absent in standard visual pretraining.

Multiplexed fluorescence microscopy improves tissue segmentation by providing complementary channels including nuclear (DAPI) and membrane (E-cadherin), that together encode richer spatial context than single-channel imaging alone. However, multiplexed models require all channels at inference, limiting deployment where only a subset is available. This work proposes a cross-modal knowledge distillation framework that transfers semantic information from a frozen foundation model teacher processing multiplexed input to a lightweight student operating on the nuclear channel only. The distillation objective combines MSE-based probability matching, boundary-aware supervision, and learnable uncertainty weighting. SAM ViT-H and CellSAM are evaluated as teachers across four U-Net students: Swin-Tiny (27M), ResNet18 (11M), EfficientNet-B0 (5.3M), and MobileNetV3 (1.5M), on TissueNet and BBBC038. On TissueNet, the SAM-distilled Swin-Tiny student achieves Dice 78.36 (plus or minus 1.44), a 13.05-point improvement over the no-KD baseline (65.31 plus or minus 1.35) and 87.9% recovery of teacher oracle performance (89.12 plus or minus 1.21) at a 23x parameter reduction. KD consistently improves all four students by approximately 12 Dice points, confirming architecture-agnostic distillation. SAM ViT-H outperforms CellSAM as teacher across all settings. Cross-dataset evaluation on BBBC038 shows consistent gains without teacher retraining.

While on-policy distillation offers dense supervision for training small reasoning models, its optimization dynamics in the multimodal domain remain under-explored. In this work, we challenge the standard monolithic view of Vision-Language Model (VLM) distillation by mathematically decomposing the loss into two distinct components: the language prior and visual grounding. Our analysis uncovers that gradient vectors for these components are nearly orthogonal, indicating that the objective of aligning with the teacher's language distribution is geometrically independent from the objective of matching its visual perception. Consequently, standard optimization passively follows a suboptimal compromise trajectory that implicitly balances the two objectives. Hypothesizing that visual grounding constitutes the primary bottleneck for vision-language reasoning, we introduce Visual Gradient Steering (VGS), a method that dynamically reorients the update vector to prioritize the visual subspace. Experimental results on multiple distillation settings and complex multimodal benchmarks demonstrate that VGS significantly outperforms the standard monolithic formulation of on-policy distillation, achieving superior grounding with minimal training overhead.

Point cloud segmentation is a fundamental task in 3D scene understanding. Its progress is constrained by the high cost and time required for dense 3D annotations, making labeled samples difficult to obtain. Beyond annotation scarcity, different sensing modalities face inherent limitations. 2D images provide rich texture and appearance cues, yet they lack explicit depth and geometric structure. In contrast, 3D point clouds capture accurate spatial geometry but are sparse and contain no texture information. As a result, relying on a single modality restricts the richness of learned representations and weakens generalization. Although recent multi-modal methods that combine 3D point clouds with 2D images have demonstrated strong performance in tasks such as classification and retrieval, they typically depend on large-scale labeled datasets and have not been fully exploited for data-efficient dense prediction. To address these limitations, we propose a novel cross-modal knowledge distillation framework, xModel-KD, for 3D point cloud segmentation. Our method exploits the complementary strengths of 2D texture and 3D geometry by learning unified per-point representations through cross-modal alignment. Specifically, we design a cross-modal fusion encoder trained with a contrastive objective that enforces feature consistency between corresponding 2D and 3D representations across multiple views. By integrating powerful pre-trained backbones with a targeted fusion strategy, the proposed framework effectively transfers appearance cues from images to geometry-aware point features. Experimental results show that cross-modal fusion achieves a 2% absolute improvement in mIoU over a LiDAR-only baseline, demonstrating the benefit of leveraging complementary multi-modal information for scalable and annotation-efficient 3D scene understanding.

When using reinforcement learning (RL) for contact-rich robotic manipulation, vision can provide task-relevant information that complements robot proprioception. However, vision-enabled policies tend to overfit to the visual conditions seen during training, limiting their robustness and transferability. We present a human-in-the-loop RL framework that employs teacher-student distillation to achieve robust performance across multiple task variants, trained entirely in the real world without requiring domain randomization or data augmentation. A vision-enabled teacher distills its knowledge into a vision-free student that relies solely on pose, twist, and wrench sensing, combining fast training with strong task generalization. On the real-world NIST assembly benchmark board, our approach achieves 95% overall success after approximately 50 minutes of training on 3 representative tasks, including robust generalization to 8 unseen task variants. Fine-tuning with distillation achieves full success on the most challenging task. We demonstrate that the resulting policies outperform baselines in both robustness and adaptability. Page: https://tuwien-asl.github.io/VE2VF/.

Unified multimodal embedding spaces have become the standard interface for cross-modal retrieval and multimodal RAG, and recent audio-video-text (AVT) encoders extend this setting to three modalities. Such encoders can produce a joint (T,V,A) embedding whenever all three modalities are available, but standard pairwise InfoNCE objectives leave this signal unused during training. We close this gap with fusion-as-teacher distillation, which treats a stop-gradient copy of the fused embedding as a teacher signal for the single-modal embeddings, paired with a Tuple-InfoNCE term that supervises the fused embedding directly. We instantiate this objective as OmniRetriever-7B. Across six zero-shot retrieval benchmarks, OmniRetriever-7B surpasses the closed-source Gemini Embedding 2 by 13.3-18.0 R@1 on Clotho and SoundDescs, and reaches the contemporary zero-shot specialist band of open video-text encoders on MSR-VTT and MSVD. To stress-test joint representations, we further release OmniRetriever-Bench, a 12-direction AVT retrieval benchmark totaling 3782 triples; on it OmniRetriever-7B attains AVG-all 34.84, improving over Gemini Embedding 2 by 1.72 and over the best prior open-source AVT method by 8.03.

Electroencephalography (EEG) visual decoding remains challenging due to the modality gap between low-SNR neural signals and highly structured vision--language spaces, making direct cross-modal alignment unstable. To address this, we propose STAMBRIDGE, a versatile two-stage framework that sequentially tackles feature conditioning and cross-modal alignment. First, we introduce a Spectral-Temporal Amplitude-aware Modulation (STAM) to extract well-conditioned EEG representations. By replacing hard frequency masking with amplitude-derived soft channel weighting and multi-scale temporal convolutions, STAM explicitly preserves frequency-aware transients while reducing the risk of time-domain ringing artifacts. Building upon these robust neural features, we further introduce a model-agnostic Mid-Feature Semantic Bridge (MFSB) that constructs a regularized intermediate space through directed cross-modal interactions, enabling staged distillation and more stable semantic alignment. Experiments on the THINGS-EEG benchmark show competitive 200-way zero-shot retrieval performance, with 34.50% Top-1 and 65.95% Top-5 accuracy. In addition, embeddings learned by STAMBRIDGE produce semantically coherent image reconstructions with a diffusion model, demonstrating robust EEG-to-vision semantic alignment. The code is available at: https://github.com/thabeatmjh/STAMBRIDGE.

The platonic representation hypothesis suggests that sufficiently large models converge to a shared representation geometry, even across modalities. Motivated by this, we ask: Can the semantic knowledge of a language model efficiently improve a vision model? As an answer, we introduce TextTeacher, a simple auxiliary objective that injects text embeddings as additional information into image classification training. TextTeacher uses readily available image captions, a pre-trained and frozen text encoder, and a lightweight projection to produce semantic anchors that efficiently guide representations during training while leaving the inference-time model unchanged. On ImageNet with standard ViT backbones, TextTeacher improves accuracy by up to +2.7 percentage points (p.p.) and yields consistent transfer gains (on average +1.0 p.p.) under the same recipe and compute. It outperforms vision knowledge distillation, yielding more accuracy at a constant compute budget or similar accuracy, but 33% faster. Our analysis indicates that TextTeacher acts as a feature-space preconditioner, shaping deeper layers in the first stages of training, and aiding generalization by supplying complementary semantic cues. TextTeacher adds negligible overhead, requires no costly multimodal training of the target model and preserves the simplicity and latency of pure vision models. Project page with code and captions: https://nauen-it.de/publications/text-teacher

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10% highest-density patches have 3.6$\times$ larger residual drift than the bottom 10% lowest-density patches and 3.5$\times$ larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer-bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models and hybrid architectures. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average with a 4.12$\times$ throughput and a 68% memory saving at 128k context, with no additional parameters and no inference-time cost.