GPQA-Diamond

The 'GPQA-Diamond' dataset/benchmark contains multi-agent debate scenarios used to evaluate the mechanisms of stance convergence and conformity in large language models (LLMs).

Papers using GPQA-Diamond (3)

Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate2026ATLAS: Agentic Test-time Learning-to-Allocate Scaling2026Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning2025