Research map / Vision-language and multimodal / Human and embodied modeling
DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
x
Problem
x
Approach
x
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DEXTER instead of DiffExplainer for visual explanation generation | SalientImageNet visual explanations · SalientImageNet | CLIP-IQA: 0.89 ± 0.09 -> 0.94 ± 0.03 | Helped |
| Use DEXTER instead of DiffExplainer for visual explanation generation | SalientImageNet visual explanations · SalientImageNet | Semantic CLIP-IQA: 0.90 ± 0.05 -> 0.96 ± 0.03 | Helped |
| Use DEXTER for slice discovery and DRO debiasing | CelebA · CelebA | worst-slice accuracy: 90.4 ± 0.9 -> 91.3 ± 0.01 | Helped |
| Use DEXTER for slice discovery and DRO debiasing | Waterbirds · Waterbirds | worst-slice accuracy: 92.4 ± 0.8 -> 90.5 ± 0.1 | Hurt |
| Use DEXTER optimized prompts instead of class-label baseline prompts | SalientImageNet activation maximization · SalientImageNet | mean activation score: 64.73 ± 31.34 -> 75.43 ± 23.17 | Helped |
| Add auxiliary mask loss to single-word prompt optimization (ablation) | DEXTER · SalientImageNet, single-word prompts | mean activation score: 23.73 ± 32.91 -> 43.93 ± 38.68 | Helped |
| Add auxiliary mask loss to multi-word prompt optimization (ablation) | DEXTER · SalientImageNet, multi-word prompts | mean activation score: 11.83 ± 23.33 -> 75.43 ± 23.17 | Helped |
| Use multi-word instead of single-word prompts with auxiliary mask loss (ablation) | DEXTER · SalientImageNet | mean activation score: 43.93 ± 38.68 -> 75.43 ± 23.17 | Helped |
| Increase the soft prompt count from P=1 to P=3 (ablation) | DEXTER · SalientImageNet | average activation score: 87.75 ± 11.68 -> 59.75 ± 32.44 | Hurt |
| Increase the soft prompt count from P=1 to P=5 (ablation) | DEXTER · SalientImageNet | average activation score: 87.75 ± 11.68 -> 73.75 ± 34.20 | Hurt |
| Increase the soft prompt count from P=1 to P=10 (ablation) | DEXTER · SalientImageNet | average activation score: 87.75 ± 11.68 -> 44.50 ± 35.59 | Hurt |
| Use multi-run top-word selection instead of frequency-ranked top k=5 words (ablation) | DEXTER slice discovery · Waterbirds | F1-score Class 1: 0.66 -> 0.76 | Helped |
Closest papers by meaning
- Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics (CVPR 2025 · similarity 0.89)
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.87)
- Comparing the Decision-Making Mechanisms by Transformers and CNNs via Explanation Methods (CVPR 2024 · similarity 0.86)
- DiVE-k: DIFFERENTIAL VISUAL REASONING FOR FINE-GRAINED IMAGE RECOGNITION (ICLR 2026 · similarity 0.86)
- The Abstraction Gap in Vision-Language Causal Reasoning (ICML 2026 · similarity 0.86)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.86)
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.