In a new paper, Anthropic researchers tested the “faithfulness” of CoT models’ reasoning by slipping them a cheat sheet and waiting to see if they acknowledged the hint. The researchers wanted to see ...