Research map / Vision-language and multimodal / Multimodal LLMs
DASH: Detection and Assessment of Systematic Hallucinations of VLMs
Maximilian Augustin, Yannic Neuhaus, Matthias Hein
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 optimized image queries (DASH-OPT) instead of text queries (DASH-LLM) (ablation) | DASH on PaliGemma · 380 object classes | total images found: 99.3K -> 221.7K | Improved |
| use optimized image queries (DASH-OPT) instead of text queries (DASH-LLM) (ablation) | DASH on PaliGemma · 380 object classes | total clusters: 1892 -> 3895 | Improved |
| use optimized image queries (DASH-OPT) instead of text queries (DASH-LLM) (ablation) | DASH on LLaVA-NeXT Vicuna · 380 object classes | total clusters: 3632 -> 4632 | Improved |
| use optimized image queries (DASH-OPT) instead of text queries (DASH-LLM) (ablation) | DASH on LLaVA-NeXT Mistral · 380 object classes | total images found: 78.5K -> 133.8K | Improved |
| use optimized DASH-OPT images instead of DASH-LLM for transfer to other VLMs (ablation) | DASH images from PaliGemma · transfer to LLaVA-NeXT Vicuna | transfer rate: 0.49 -> 0.43 | Hurt |
| swap LLaVA-NeXT LLM backbone from Llama 3.0 to Vicuna | LLaVA-NeXT · transfer of PaliGemma DASH-LLM images | transfer rate: 0.26 -> 0.49 | Hurt |
| swap Prismatic vision encoder from SigLIP to CLIP | Prismatic Vicuna · DASH image transfer | average transfer rate: 0.43 -> 0.60 | Hurt |
| scale Qwen2-VL language model from 7B to 72B | Qwen2-VL · DASH image transfer | average transfer rate: 0.19 -> 0.18 | No clear effect |
| fine-tune PaliGemma on DASH images | PaliGemma · DASH-B | accuracy: 56.4% -> 68.0% | Improved |
| fine-tune PaliGemma on DASH images | PaliGemma · Amber Existence | accuracy: 93.2% -> 95.4% | Improved |
| fine-tune PaliGemma on DASH images | PaliGemma · R-Bench | accuracy: 79.9% -> 80.2% | No clear effect |
| fine-tune PaliGemma on DASH images | PaliGemma · POPE | accuracy: 87.2% -> 86.4% | Hurt |
Closest papers by meaning
- Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention (AAAI 2026 · similarity 0.92)
- Analyzing and Mitigating Object Hallucination in Large Vision-Language Models (ICLR 2024 · similarity 0.92)
- Multi-Object Hallucination in Vision Language Models (NeurIPS 2024 · similarity 0.91)
- Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations (ICLR 2025 · similarity 0.91)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.91)
- Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention (CVPR 2025 · similarity 0.91)
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.
