consistency-evaluation

This paper proposes a framework to test whether LLM estimates obey statistical self-consistency (law of total probability) across subpopulations, finding widespread violations and the 'macro fallacy' where fine-grained estimates align better with human data.