Agents-as-Peers: Bias-Controlled LLM Evaluation as Multi-Agent Peer Assessment
Agents-as-Peers is a distributed evaluation protocol where LLMs autonomously generate tasks, answer them using live web grounding, and then judge each other's answers. This multi-agent peer assessment framework produces stable rankings that correlate with accuracy on standard benchmarks like TruthfulQA and GSM8K. It demonstrates that peer evaluation can be a scalable alternative to human benchmarks when biases are explicitly measured.