AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features

Chancharik Mitra, Brandon Huang, Tianning Chai, Zhiqiu Lin, Assaf Arbelle, Rogerio Feris, Leonid Karlinsky, Trevor Darrell, Deva Ramanan, Roei Herzig

ICCV 2025 · Vision-language and multimodal › CLIP and zero-shot · approach: Self-distillation for vision-language pretraining · 16 citations (OpenAlex)

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
use class centroid classification instead of KNN for head selection (ablation)SAVs (LLaVA-OneVision-7B) · MHaluBenchaccuracy: 53.0 -> 80.8Improved
use class centroid instead of linear probing (ablation)SAVs (LLaVA-OneVision-7B) · MHaluBench / NaturalBench / EuroSATaccuracy: MHB 82.5 vs 80.8; NB 32.9 vs 35.1; ES 83.1 vs 86.7Mixed
select sparse heads instead of sparse layers (ablation)SAVs (LLaVA-OneVision-7B) · NaturalBench (group score)accuracy: 28.4 -> 35.1Improved
select sparse heads instead of sparse layers (ablation)SAVs (LLaVA-OneVision-7B) · EuroSATaccuracy: 81.8 -> 86.7Improved
add 2 or 5 noisy examples per class (ablation)SAVs (LLaVA-OneVision-7B) · NaturalBench (group score)accuracy: 35.1 -> 36.1No clear effect
add 10 noisy examples per class out of 20 (ablation)SAVs (LLaVA-OneVision-7B) · MHaluBenchaccuracy: 82.0 -> 50.3Hurt
extract from last token instead of first token (ablation)SAVs (LLaVA-OneVision-7B) · EuroSATaccuracy: 24.9 -> 86.7Improved
extract from last token instead of middle token (ablation)SAVs (LLaVA-OneVision-7B) · NaturalBench (group score)accuracy: 2.4 -> 35.1Improved
add online learning with randomized weighted voting (ablation)SAVs (LLaVA-OneVision-7B) · MHaluBenchaccuracy: 82.0 -> 73.2Hurt
use 20 one-shot examples instead of formatted ICL input (ablation)SAVs (LLaVA-OneVision-7B) · MHaluBenchaccuracy: 28.3 -> 82.0Improved
add SAVs to the zero-shot LMMSAVs (LLaVA-OneVision-7B) · VLGuardaccuracy: 31.4 -> 94.3Improved
add 4-shot ICL to the zero-shot LMMSAVs (LLaVA-OneVision-7B) · MHaluBenchaccuracy: 34.7 -> 25.0Hurt

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