Evaluating Language-Model Social Media Agents under Realistic Interaction and Platform Constraints

Abstract

Llm Social Agents is advancing through efforts to align performance with evidence quality, resource limits, and transfer across settings. The present synthesis investigates moving agent evaluation from isolated prompts to persistent identities, social feedback, and platform dynamics. The comparison integrates 1 focal paper with 11 independently retrieved publications whose authorship and venue fields were independently checked. The analysis is organized around behavioral realism, memory, interaction effects, safety, and benchmark validity. The comparison does not regard outcomes from different settings as equivalent, the review compares problem definitions, methodological assumptions, and validation boundaries. Across the literature, the evidence indicates that advances in LLM social agents become credible when representation, objective, and evaluation protocol are evaluated together and when uncertainty about distribution shift is reported explicitly. The resulting framework links method selection to downstream risk and the conditions that weaken external validity, and proposes a research agenda centered on explicit comparators, boundary tests, and reproducible artifacts.