PSI-VQA: Evaluating Multimodal LLM Social Reasoning under Ambiguous Pedestrian Intent in Automated Driving

• Paper

Summary

PSI-VQA evaluates multimodal language models on ambiguous pedestrian intent in automated driving. Its 1,475 questions cover 244 dashcam videos and ask models to predict crossing behavior, identify supporting and opposing visual cues, and locate decisive moments. The benchmark extends annotations from the PSI dataset and examines how well models reason about fine-grained social and temporal evidence.

Citation

@inproceedings{Wang26PSIVQA,
  author = {Shaozhi Wang and Yin-Chun Lu and Zheng Tang and Han Zhang and David C. Anastasiu and Renran Tian},
  title = { {PSI-VQA}: {E}valuating multimodal {LLM} social reasoning under ambiguous pedestrian intent in automated driving},
  booktitle = {Proc. ECCV Workshops},
  address = {Malm{\"o}, Sweden},
  year = {2026}
}