AI Paper Atlas

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

ICLR 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives

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 changedWhereMeasuredEffect
add visual attention alignment (VAT) (ablation)CompoDistill w/o TAF · Compositional reasoning avgaccuracy: 62.9 -> 65.0Improved
add Teacher Adapter Fetch (TAF) (ablation)CompoDistill w/o VAT · VQA and CR avgaccuracy: VQA 56.8 -> 61.3; CR 62.9 -> 63.8Improved
add VAT on top of TAF (ablation)CompoDistill · Compositional reasoning avgaccuracy: 63.8 -> 66.7Improved
add VAT alone (no TAF) (ablation)student KD baseline · GQAaccuracy: 57.8 -> 54.3Hurt
cosine similarity attention loss instead of MSE (ablation)CompoDistill · Compositional reasoning avgaccuracy: 65.2 -> 66.7Improved
MSE or KL attention loss vs none (ablation)CompoDistill w/o attention loss · VQA avgaccuracy: 61.3 -> 60.3 (MSE), 60.7 (KL)Hurt
distill attention in intermediate layers rather than early or later (ablation)CompoDistill · Compositional reasoning avgaccuracy: 63.7 (early), 64.6 (later) -> 66.7Improved
intermediate layers vs all layers (ablation)CompoDistill · Compositional reasoning avgaccuracy: 66.6 -> 66.7No clear effect
group layer matching vs simple/adaptive matching (ablation)CompoDistill · Compositional reasoning avgaccuracy: 65.6 (simple), 65.7 (adaptive) -> 66.7Improved
final SFT stage after DFT (ablation)CompoDistill three-stage · VQA avgaccuracy: 21.6 -> 62.9Improved
double training data to 2.4MCompoDistill · Compositional reasoning avgaccuracy: 66.7 -> 69.9Improved
larger teacher (Qwen2.5-7B vs Qwen1.5-4B)CompoDistill 1.8B student · Compositional reasoning avgaccuracy: 66.7 -> 67.8Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy