Research map / Label-efficient and robust learning / Knowledge distillation
CompoDistill: Attention Distillation for Compositional Reasoning in Multimodal LLMs
Jiwan Kim, Kibum Kim, Sangwoo Seo, Chanyoung Park
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 visual attention alignment (VAT) (ablation) | CompoDistill w/o TAF · Compositional reasoning avg | accuracy: 62.9 -> 65.0 | Improved |
| add Teacher Adapter Fetch (TAF) (ablation) | CompoDistill w/o VAT · VQA and CR avg | accuracy: VQA 56.8 -> 61.3; CR 62.9 -> 63.8 | Improved |
| add VAT on top of TAF (ablation) | CompoDistill · Compositional reasoning avg | accuracy: 63.8 -> 66.7 | Improved |
| add VAT alone (no TAF) (ablation) | student KD baseline · GQA | accuracy: 57.8 -> 54.3 | Hurt |
| cosine similarity attention loss instead of MSE (ablation) | CompoDistill · Compositional reasoning avg | accuracy: 65.2 -> 66.7 | Improved |
| MSE or KL attention loss vs none (ablation) | CompoDistill w/o attention loss · VQA avg | accuracy: 61.3 -> 60.3 (MSE), 60.7 (KL) | Hurt |
| distill attention in intermediate layers rather than early or later (ablation) | CompoDistill · Compositional reasoning avg | accuracy: 63.7 (early), 64.6 (later) -> 66.7 | Improved |
| intermediate layers vs all layers (ablation) | CompoDistill · Compositional reasoning avg | accuracy: 66.6 -> 66.7 | No clear effect |
| group layer matching vs simple/adaptive matching (ablation) | CompoDistill · Compositional reasoning avg | accuracy: 65.6 (simple), 65.7 (adaptive) -> 66.7 | Improved |
| final SFT stage after DFT (ablation) | CompoDistill three-stage · VQA avg | accuracy: 21.6 -> 62.9 | Improved |
| double training data to 2.4M | CompoDistill · Compositional reasoning avg | accuracy: 66.7 -> 69.9 | Improved |
| larger teacher (Qwen2.5-7B vs Qwen1.5-4B) | CompoDistill 1.8B student · Compositional reasoning avg | accuracy: 66.7 -> 67.8 | Improved |
Closest papers by meaning
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.89)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)
- CoMT: A Novel Benchmark for Chain of Multi-modal Thought on Large Vision-Language Models (AAAI 2025 · similarity 0.89)
- VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models (ICLR 2026 · similarity 0.88)
- VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.88)
- Toward Cognitive Supersensing in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.88)
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.
