Can LLMs Predict Academic Collaboration? Topology Heuristics vs. LLM-Based Link Prediction on Real Co-authorship Networks

This collection highlights two security- and privacy-relevant findings: (1) Large language models (Qwen2.5-72B-Instruct) can predict future academic collaborations from author metadata alone (AUROC 0.714–0.789 for new-edge prediction; AUROC 0.652 even when authors share no common neighbor), relying primarily on research-concept metadata; providing pre-computed graph features can degrade LLM performance via anchoring. (2) AgentSocialBench, a new benchmark for human-centered agentic social networks, demonstrates systemic privacy failures in multi-agent settings: cross-domain and cross-user agent