Affective computing has moved from categorical emotion recognition to open‑ended analysis using large multimodal models. Yet affective science treats emotion as a dynamic episode shaped by appraisal, regulation and social interpretation, which remains under‑explored computationally. We introduce TRACE, a cognition‑oriented framework that formalizes an affective episode into three interrelated stages—Condition, Affect, and Effect—and integrates observable cues with cognitive factors such as internal stance and regulation of emotional display. Based on this formulation, TRACE‑Bench evaluates multimodal models in real‑world social scenes through five tasks: grounded affect recognition, regulation decoding, cause reasoning, effect reasoning, and full‑chain reconstruction, comprising 3,746 structured QA pairs over 646 videos. A human‑model comparison reveals a substantial performance gap, and emotion‑specialized models generally lag behind general‑purpose multimodal LLMs. Common failures include treating displayed behavior as genuine feeling and fabricating unsupported events in long‑chain generation. To address this, we propose TRACER, a cognition‑grounded structured reasoning method that couples each inference with explicit premises drawn from factual observations, cognitive appraisals, and upstream conclusions, forming a traceable graph of intermediate and target conclusions. Experiments show TRACER outperforms all baselines on each of the five tasks. Project page: https://cogaffc.github.io/TRACE
Review