The Social World Model Evaluation Gap

An audit of 113 social world model evaluations published from 1999 to 2026 finds that only 2 of 9 primary evaluations directly measure a real person's response to a system-selected action, and none of 25 checked language-model works meet the proposed evaluation conditions. The study proposes a 1-step negotiation experiment to separate response prediction from other evaluation approaches, noting the findings describe the audited sample rather than field-wide prevalence.