Announcement
AI agents struggle to read the room, NormGym researchers report
NormGym’s creators say agents learn more from their own penalties than from watching others get penalized.
TLDR
In NormGym’s tests, the team reports, Claude Opus 5 led tested LLM agents at 55.3% success on familiar layouts, counting only episodes where agents finished tasks and obeyed hidden community rules. Giving GPT-5.6 Sol the rule raised its score from 44.0% to 90.7%. The team’s diagnosis: figuring out the rule is harder than following it.
Combined views
4.7K
6 Sources, first seen ago

Comments