Why Models Lack Confidence: Over-Alignment Kills Creativity

ctjlewis · x · 2026-08-11

The discussion points out that current LLMs are often trained to lack self-belief, leaning towards conservative answers like 'this is an open problem I can't solve.' Users can nudge the model out of this distribution of evasive responses to increase the chances of successfully solving complex problems.

Related event: Over-Alignment Causes LLMs to Lack Confidence and Self-Censor(3 posts)→

Original post →

More from Models

Models channel →