Benchmarking Open-Weight Foundation Models for Global AI Technical Governance

Collection of new AI governance/measurement papers (arXiv, 26 Jun 2026) addressing systematic evaluation and societal risks of generative models. Key contributions: (1) Benchmarking open-weight frontier LLMs against the Global AI Dataset v2 (GAID v2) to quantify geographic bias using ~2,990 country-metric-year observations across 227 countries and a five-category response taxonomy (verified accuracy, hallucinated fabrication, honest refusal, qualitative hedging, misattribution); (2) Demonstration that LLM patient simulations (eating‑disorder personas) are overly stable but systematically over‑