Research map / Label-efficient and robust learning / Knowledge distillation
Module-wise Adaptive Distillation for Multimodality Foundation Models
Chen Liang, Jiahui Yu, Ming-Hsuan Yang, Matthew Brown, Yin Cui, Tuo Zhao, Boqing Gong, Tianyi Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add OPTIMA module-adaptive distillation to layerwise distillation (CoCa-Tiny12) | CoCa-Tiny12 · COCO Caption | CIDEr: 116.8 -> 121.2 | Improved |
| Add OPTIMA (CoCa-Tiny12) | CoCa-Tiny12 · NLVR2 test-p | accuracy: 80.4 -> 81.3 | Improved |
| Add OPTIMA (CoCa-Tiny12) | CoCa-Tiny12 · VQA test-std | accuracy: 71.2 -> 71.6 | Improved |
| Add OPTIMA (CoCa-Tiny12) | CoCa-Tiny12 · SNLI-VE test | accuracy: 83.9 -> 84.3 | Improved |
| Add OPTIMA (CoCa-Tiny6) | CoCa-Tiny6 · COCO Caption | CIDEr: 112.2 -> 113.5 | Improved |
| Add OPTIMA (CoCa-Tiny6) | CoCa-Tiny6 · NLVR2 test-p | accuracy: 76.5 -> 77.0 | Improved |
| Distill fixed arm Img+Txt+Multi instead of Img+Multi (ablation) | CoCa-Tiny · NLVR2 | accuracy: 81.1 -> 80.4 | Hurt |
| Replace fixed-arm and random-arm distillation with OPTIMA arm selection (ablation) | CoCa-Tiny · multimodal tasks | student performance | Improved |
| Use OPTIMA instead of layerwise distillation | CoCa-Tiny · dev set | prediction loss and KL discrepancy | Improved |
| Reward based only on KD loss (r_KD) instead of combined reward (ablation) | OPTIMA · NLVR2 | accuracy | Mixed |
| Reward based only on layerwise losses (r_LWD) instead of combined reward (ablation) | OPTIMA · COCO Caption | CIDEr | Mixed |
| Vary EMA halflife of reward distribution too large or too small (ablation) | OPTIMA · student performance | accuracy | Hurt |
Closest papers by meaning
- Scalable Model Merging with Progressive Layer-wise Distillation (ICML 2025 · similarity 0.87)
- Your Student is Better Than Expected: Adaptive Teacher-Student Collaboration for Text-Conditional Diffusion Models (CVPR 2024 · similarity 0.86)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.86)
- Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling (ICLR 2026 · similarity 0.85)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.85)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
