Research map / Vision-language and multimodal / Multimodal LLMs
Entropy-Gradient Grounding: Training-Free Evidence Retrieval in Vision-Language Models
Marcel Gröpl, Jaewoo Jung, Seungryong Kim, Marc Pollefeys, Sunghwan Hong
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 |
|---|---|---|---|
| use spatial-entropy stopping instead of confidence stopping (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 70.16 -> 73.3 | Improved |
| use spatial-entropy stopping instead of 1 fixed iteration (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 60.73 -> 73.3 | Improved |
| use spatial-entropy stopping instead of 3 fixed iterations (ablation) | Ours on LLaVA-1.6 · DocVQA | accuracy: 61.80 -> 65.07 | Improved |
| add iterative refinement with stopping (4.80 s) versus single pass (ablation) | Ours on LLaVA-1.6 · RWQA | inference time: 2.22 s -> 4.80 s | Hurt |
| backpropagate Shannon entropy instead of max-probability log (ablation) | Ours on LLaVA-1.6 · DocVQA | accuracy: 64.53 -> 65.07 | Improved |
| use top-P entropy instead of full entropy (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 73.30 -> 72.25 | No clear effect |
| select 2 regions instead of 1 (ablation) | Ours on LLaVA-1.6 · DocVQA | accuracy: 61.25 -> 65.07 | Improved |
| select 2 regions instead of 1 (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 69.63 -> 73.30 | Improved |
| increase regions from 2 to 4 (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 73.30 -> 72.25 | No clear effect |
| compute gradient at last layer 32 instead of layer 15 (ablation) | Ours on LLaVA-1.6 · DocVQA | accuracy: 50.99 -> 65.07 | Improved |
| compute gradient at layer 20 instead of last layer (ablation) | Ours on LLaVA-1.6 · V* | accuracy: 73.30 -> 78.01 | Mixed |
| backpropagate from first token instead of second (ablation) | Ours on LLaVA-1.6 · TextVQA | accuracy: 65.27 -> 67.96 | Improved |
Closest papers by meaning
- Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.89)
- DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.88)
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.87)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.87)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.87)
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.
