Research map / Vision-language and multimodal / Multimodal LLMs
TruthLens: Object Hallucination Detection via Self-Evaluating Truthfulness Scores in LVLMs
Yanqi Wu, Runhe Lai, Xinhua Lu, Qichao Chen, Zhiping Zhou, Jia-Xin Zhuang, Weijiang Yu, Ruixuan Wang
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 |
|---|---|---|---|
| fine-tune with MSE truthfulness objective plus KL (ablation) | TruthLens · MSCOCO | AUROC: 34.80 -> 89.02 | Improved |
| add class-level loss reweighting (ablation) | TruthLens · MSCOCO | AUROC: 89.02 -> 90.57 (LLaVA-1.5-7B); 85.69 -> 91.46 (Qwen2.5-VL-7B) | Improved |
| keep KL divergence constraint (ablation) | TruthLens · MSCOCO | generation quality | Improved |
| swap special token (<|image_pad|> vs vision_start/end) (ablation) | TruthLens · MSCOCO | AUROC: 91.20-91.54 on Qwen2.5-VL-7B | No clear effect |
| supervise first vs last vs averaged sub-token (ablation) | TruthLens · MSCOCO | AUROC | No clear effect |
| increase number of sampled rollouts (ablation) | TruthLens · MSCOCO | AUROC | Improved |
| use TruthLens fine-tuning on LLaVA-1.5-7B | LLaVA-1.5-7B · 13 general benchmarks (e.g. POPE) | accuracy: POPE 85.76 -> 85.67; MME 1468.56 -> 1474.59 | No clear effect |
| replace best baseline with TruthLens | LLaVA-OneVision-1.5-8B · MSCOCO | AUROC: 76.94 -> 93.04 | Improved |
| train on MSCOCO only, test on Object365 | TruthLens · Object365 | AUROC: 72.59 -> 82.14 (LLaVA-OV-1.5-8B vs SVAR) | Improved |
| apply TruthLens-guided caption revision | LLaVA-1.5-7B · MSCOCO captions | CHAIR_s: 49.40 -> 43.78 | Improved |
| apply TruthLens to weak-reasoning models on CLEVR | LLaVA-1.5 · CLEVR | AUROC: 59.89 -> 63.34 (7B) | Mixed |
Closest papers by meaning
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.91)
- Analyzing and Mitigating Object Hallucination in Large Vision-Language Models (ICLR 2024 · similarity 0.91)
- Instruction Lens Score: Your Instruction Contributes a Powerful Object Hallucination Detector for Multimodal Large Language Models (ICML 2026 · similarity 0.91)
- Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.90)
- Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention (AAAI 2026 · similarity 0.90)
- Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024 · similarity 0.90)
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.
