Cross-Modal Consistency and Safety Evaluation in Generative Agent Social Simulations
ACL 2026
X-CASE is a benchmark dataset of 1,000 multimodal social activity scenarios designed
to evaluate how generative AI agents detect and correct unsafe behaviour during iterative
plan revision. Every scenario contains:
A natural-language social activity description (e.g. a beach party, rooftop gathering).
An unsafe hourly plan — 11 activities from 7 PM to 5 AM that include… See the full description on the dataset page:
https://huggingface.co/datasets/adonaivera/X-CASE.
