Research map / Vision-language and multimodal / Multimodal LLMs
Hallucinatory Image Tokens: A Training-free EAZY Approach to Detecting and Mitigating Object Hallucinations in LVLMs
Liwei Che, Tony Qingze Liu, Jing Jia, Weiyi Qin, Ruixiang Tang, Vladimir Pavlovic
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 iterative per-object hallucination detection (ablation) | EAZY (LLaVA-1.5) · CHAIR | CHAIR_S: 38.8 -> 37.8 | Improved |
| use iterative per-object detection (ablation) | EAZY (Shikra) · CHAIR | CHAIR_S: 26.6 -> 26.2 | Improved |
| use iterative per-object detection (ablation) | EAZY (LLaVA-NEXT) · CHAIR | CHAIR_I: 8.3 -> 8.5 | Hurt |
| use layer 15 attention instead of layer 10 (ablation) | EAZY (LLaVA-1.5) · CHAIR | CHAIR_S: 39.8 -> 38.8 | Improved |
| use layer 15 attention instead of layer 25 (ablation) | EAZY (LLaVA-1.5) · CHAIR | CHAIR_S: 41.7 -> 38.8 | Improved |
| zero out at input token space rather than ViT output (ablation) | EAZY (LLaVA-1.5) · CHAIR | CHAIR_S: 40.4 -> 38.8 | Improved |
| zero out at image patches rather than input tokens (ablation) | EAZY (LLaVA-1.5) · CHAIR | CHAIR_S: 38.8 -> 39.0 | No clear effect |
| zero-out replacement instead of average embedding (ablation) | top-K HIT processing · Hall-COCO | images without OH: 55.17% -> 68.27% | Improved |
| zero-out replacement instead of random token (ablation) | top-K HIT processing · Hall-COCO | removed OH: 63.33% -> 74.17% | Improved |
| zero-out replacement instead of neighbor token (ablation) | top-K HIT processing · Hall-COCO | removed OH: 61.25% -> 74.17% | Improved |
| zero out top-5 candidate HITs | LLaVA-1.5 · Hall-COCO | images with hallucination: 100% -> 31.73% | Improved |
| use EAZY over best baseline IC for detection | LLaVA-1.5 · Hall-COCO | accuracy: 62.56 -> 78.77 | Improved |
Closest papers by meaning
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.94)
- One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination (CVPR 2026 · similarity 0.93)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.93)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.93)
- Image Token Matters: Mitigating Hallucination in Discrete Tokenizer-based Large Vision-Language Models via Latent Editing (NeurIPS 2025 · similarity 0.93)
- Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations (ICLR 2025 · similarity 0.92)
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.
