AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Attention Prompting on Image for Large Vision-Language Models

Runpeng Yu, Weihao Yu, Xinchao Wang

ECCV 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 29 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
overlay query-guided attention heatmap (API) on imageLLaVA-1.5 · MM-VetGPT-eval score: 32.8 -> 36.6Improved
overlay API heatmapGemini · VisWizaccuracy: 50.28 -> 58.58Improved
overlay API heatmapGPT-4V · MM-VetGPT-eval score: 67.00 -> 67.70No clear effect
add 'Let's think step by step' text promptLVLMs · VQA benchmarksaccuracy/score: Gemini MM-Vet 59.0 -> 30.6Hurt
use query-agnostic visual prompts (FGVP, SoM)LLaVA-1.5 · TextVQAaccuracy: 48.32 -> 18.81 (SoM)Hurt
scale up auxiliary model (CLIP-ViT-B to ViT-L-336) (ablation)API · MMMUaccuracy: 36.03 -> 37.52Improved
weaker auxiliary CLIP model (ablation)API · MMEaccuracy: 85.50 -> 83.50Hurt
add mean filter (kernel 3) to heatmap (ablation)API · MMEaccuracy: 83.70 -> 86.20Improved
extract LLaVA attention from last two layers instead of layer 20 (ablation)API (LLaVA auxiliary) · MMEaccuracy: 86.34 -> 83.10Hurt
add reflection via evaluation (judge if highlighted region suffices, else use original image) (ablation)API (LLaVA) · LLaVA-BenchGPT-eval score: 74.80 -> 76.10Improved
add re-emphasize hint that answer lies in highlighted areas (ablation)API (LLaVA) · LLaVA-BenchGPT-eval score: 74.80 -> 72.70Hurt
ensemble API heatmaps from CLIP and LLaVA as two images (ablation)GPT-4V · LLaVA-BenchGPT-eval score: 103.60 -> 104.80Improved

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