Answers are grounded in this paper and cite their sources.
Group Emotion Analysis (GEA) is moving from closed-set group emotion recognition toward open-world understanding of multi-person social affect. Large Language Models (LLMs), and multimodal LLMs (MLLMs) in particular, are driving this shift because they return more than a label or a score: rationales, role claims, profiles, pseudo-labels, and tool-grounded reports. Existing GEA surveys organize the field by architecture or modality, which cannot capture the fact that a single model may perform multiple functions within a single pipeline. We therefore organize LLM use in GEA by functional role. Following a PRISMA-guided, evidence-layered protocol, we reviewed 105 studies from January 2020 to May 2026, separating 45 core GEA studies from 38 adjacent transferable and 22 reliability-supporting studies, so that validated group-level findings are not confused with results transferable only by analogy. We derive a taxonomy of five functional roles: Predictor, Perceiver, Generator, Data-Engine, and Hybrid. We strictly define Hybrid as tool-augmented reasoning over structured intermediate evidence rather than ordinary preprocessing. We map each role to its GEA tasks, modalities, emotion representations, datasets, and evaluation requirements. Synthesizing reported results, we find that on closed-set accuracy zero-shot systems trail trained baselines and that apparent gains often reflect permissive metrics. The value these models promise lies in their wider output space, which current benchmarks cannot yet score, so the field faces an evaluation gap rather than demonstrated parity. Because each role carries its own evidence, metrics, reliability checks, and governance concerns, we close with a role-specific research agenda.
No takes yet. Share an insight, caveat, or question.
MENSAH et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: