Werewolf benchmark shows GPT models miss social cues in deception
AIResearchers from Dongxi NLP and collaborators built a Werewolf game benchmark to test whether multimodal large models can detect deception from video, audio, and text. The source text provided is only a short lead, so no benchmark scores, model versions, or findings are confirmed here.
Source details
AI Era · AIERA · aiera.com.cnPublished · added here
Original title: EMNLP’26｜GPT也不会「读空气」，狼人杀基准测出社交盲区