COSI-Lab Dataset Models Multi-Perspective Social Intentions in Conferences

Understanding human social intention is crucial for developing truly intelligent AI systems, but existing approaches often struggle with the subjective and multi-faceted nature of human perception. The COSI-Lab project introduces a unique dataset designed to address this challenge. It captures rich, multimodal, and multi-sensor data from 32 academics interacting in a weakly scripted, real-world scientific workshop environment, including mingling sessions with genuine professional and social stakes. The core innovation of COSI-Lab is its focus on modeling "Apparent Intent Inference" (AII) from the perspective of ex-situ observers, recognizing that intentions are independent of manifest outcomes. The dataset features a novel annotation process that accounts for perceivers' interpretive tendencies, providing quantitative and qualitative analyses of intent narratives. It also offers benchmark tasks for AII, speech quality audio, privacy-preserving multimodal data, and links self-reported participant goals with annotated AII, paving the way for AI systems that can better understand and navigate complex social dynamics.