@xuanalogue on Backlist

Reasoning models are bad at judging bad reasoning (x.com)

Models trained to generate convincing chains of thought can still accept invalid reasoning, separating fluency from verification