# DARK: Diagonal-Anchored Repulsive Knowledge Distillation for Vision-Language Models under Extreme Compression

arXiv: https://arxiv.org/abs/2603.05421 · HTML: https://mlsanity.com/paper/2603.05421/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 25 of its 36 references that are in the map.
- First posted on arXiv 5 Mar 2026.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)
- Knowledge distillation / [Multimodal distillation](https://mlsanity.com/distill/multimodal-distillation/)

## Builds on (most cited first, 25 of 25)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Sigmoid Loss for Language Image Pre-Training](https://mlsanity.com/paper/2303.15343/) (2023)
- [Barlow Twins: Self-Supervised Learning via Redundancy Reduction](https://mlsanity.com/paper/2103.03230/) (2021)
- [FitNets: Hints for Thin Deep Nets](https://mlsanity.com/paper/1412.6550/) (2014)
- [Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere](https://mlsanity.com/paper/2005.10242/) (2020)
- [Reproducible scaling laws for contrastive language-image learning](https://mlsanity.com/paper/2212.07143/) (2022)
- [Relational Knowledge Distillation](https://mlsanity.com/paper/1904.05068/) (2019)
- [Contrastive Representation Distillation](https://mlsanity.com/paper/1910.10699/) (2019)
- [Improved Knowledge Distillation via Teacher Assistant](https://mlsanity.com/paper/1902.03393/) (2019)
- [MedCLIP: Contrastive Learning from Unpaired Medical Images and Text](https://mlsanity.com/paper/2210.10163/) (2022)
- [Born Again Neural Networks](https://mlsanity.com/paper/1805.04770/) (2018)
- [Regularizing Neural Networks by Penalizing Confident Output Distributions](https://mlsanity.com/paper/1701.06548/) (2017)
- [Decoupled Knowledge Distillation](https://mlsanity.com/paper/2203.08679/) (2022)
- [On the Efficacy of Knowledge Distillation](https://mlsanity.com/paper/1910.01348/) (2019)
- [Knowledge Distillation from A Stronger Teacher](https://mlsanity.com/paper/2205.10536/) (2022)
- [FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization](https://mlsanity.com/paper/2303.14189/) (2023)
- [Logit Standardization in Knowledge Distillation](https://mlsanity.com/paper/2403.01427/) (2024)
- [MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training](https://mlsanity.com/paper/2311.17049/) (2023)
- [TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance](https://mlsanity.com/paper/2309.12314/) (2023)
- [CLIP-KD: An Empirical Study of CLIP Model Distillation](https://mlsanity.com/paper/2307.12732/) (2023)
- [SonoNet: Real-Time Detection and Localisation of Fetal Standard Scan Planes in Freehand Ultrasound](https://mlsanity.com/paper/1612.05601/) (2016)
- [UniMed-CLIP: Towards a Unified Image-Text Pretraining Paradigm for Diverse Medical Imaging Modalities](https://mlsanity.com/paper/2412.10372/) (2024)
- [FetalCLIP: A Visual-Language Foundation Model for Fetal Ultrasound Image Analysis](https://mlsanity.com/paper/2502.14807/) (2025)
- [MobileCLIP2: Improving Multi-Modal Reinforced Training](https://mlsanity.com/paper/2508.20691/) (2025)

## Abstract (arXiv)

Compressing vision-language models for on-device deployment is increasingly important in clinical settings, but knowledge distillation (KD) degrades sharply when the teacher-student capacity gap spans an order of magnitude or more. We argue that, under such gaps, strict imitation of the teacher is a poor objective: much of the teacher's pairwise similarity structure reflects its own architectural biases rather than information a compact student can efficiently represent. We propose Diagonal-Anchored Repulsive Knowledge Distillation (DARK), a contrastive KD framework that decomposes the distillation loss into a diagonal term (matched image-text pairs) and an off-diagonal term (non-target similarities). The diagonal term anchors matched-pair alignment throughout training; the off-diagonal term is annealed from positive to negative weighting, transitioning the student from imitating to repelling the teacher's non-target similarity structure. We instantiate DARK by distilling FetalCLIP, a 427M-parameter fetal ultrasound vision-language model, into MobileFetalCLIP, a 75M-parameter student model with a 26× smaller visual encoder, running in 1.6,ms on an iPhone~16~Pro. The student matches or exceeds its teacher on three zero-shot benchmarks, including HC18 biometry validity (88.6% vs. 83.5%) and brain sub-plane F1 (0.784 vs. 0.702). Embedding-geometry and logit analyses show that DARK induces structured decorrelation: the student preserves teacher-aligned per-image confidence while diverging from inherited inter-class confusion, suggesting that controlled repulsion can be more efficient than imitation under extreme compression.