CrucibleBench

CrucibleBench, a proof-of-concept evaluation framework that places large language models in a persistent MUD (multi-user dungeon) over 50 turns with hidden social objectives, found that a single LLM-j…