Reproduce “AgentSocialBench: Evaluating Privacy Risks in Human-Centered Agentic Social Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Prince Zizhuang Wang, Shuli Jiang

AgentSocialBench evaluates privacy in agent networks with 352 synthetic scenarios across seven dyadic and multi-party categories and eight LLM backbones. It compares guidance levels, with the strongest combining three defenses. From no guidance to full guidance, dyadic leakage falls from 0.36 to 0.32, while multi-party leakage rises from 0.11 to 0.13 as abstraction scores increase. Results are limited to synthetic tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.