[memo]VIDHAL: Benchmarking Temporal Hallucinations in Vision LLMs

Vidhal is constructed by bootstrapping

Contrastive decoding

Attention calibration

MMHalbench

AMBER

expand beyond object-based evaluations

Contribution

A benchmark dataset dedicated to video-based hallucination evaluation of VLLMs

Create novel evaluation task of caption ordering

Hallucination

Conclusion

一旦飛ばします