TRACE-Bench frames affective inference as a multi-stage reasoning problem, not only a matter of labeling visible emotion. TRACE-Bench evaluates five linked tasks; grounded affect recognition, regulation decoding, cause reasoning, effect reasoning, and full-chain reconstruction.
The benchmark contains 3,746 structured question-answer pairs drawn from 646 multimodal social-video clips. Its design follows an affective chain connecting conditions, affect, regulation, and consequences across social interactions.
On the reported benchmark, TRACER outperformed the evaluated model baselines across all five tasks. The matched human-model comparison remained limited; humans averaged 79.15, GPT-5 thinking 58.75, and TRACER 63.58 on 50 questions per task answered by two human participants.
The results also expose material error patterns and evaluation limits. Models often inferred active regulation when the reference tactic was None, producing false-alarm rates of 76% to 96% across the named models. In a separate factual-drift diagnostic, TRACER reduced the aggregate fabrication rate from 11.0% to 4.7%, but that measure covered the specified statements and audit setup.
The proposed method separates factual observation extraction, cognition-grounded affective analysis, and structured reasoning, linking inference steps to explicit premises. These findings support evaluation controls for evidence traceability, benchmark scope, uncertainty, and unsupported-content testing; they do not establish deployment readiness, production safety, or generalization beyond the reported conditions.