Reproduce “Visual hallucination detection in large vision-language models via evidential conflict” on OpenResearch

About this paper

Tao Huang, Zhekun Liu, Rui Wang, Yang Zhang, Liping Jing

Researchers from Beijing Jiaotong University introduced a comprehensive benchmark, PRE-HAL, to evaluate visual hallucinations in Large Vision-Language Models across perception and reasoning tasks, identifying critical vulnerabilities in reasoning. They also proposed an efficient, inference-phase method based on evidential conflict, which improved hallucination detection performance, achieving average AUROC gains of 4-10% over internal baselines.

Key insights

• LVLMs exhibit substantially higher hallucination rates in complex reasoning tasks, especially relation reasoning (up to 50%), compared to perception-based tasks.
• Quantifying evidential conflict from high-level model features provides a robust and efficient internal metric for hallucination detection, effectively overcoming the calibration limitations of probability-based uncertainty metrics.
• The proposed evidential conflict method (κ_max) demonstrates stable and superior performance across varying inference temperatures and different LVLM sizes.

Summary by alphaXiv.