Suppressing 'wait/maybe' tokens reportedly boosts Qwen's accuracy in Reddit test

yangyi · x · 2026-09-28

A Reddit user ran a test set on Qwen and found that downweighting hedge words like "wait / maybe / perhaps" actually improved the model's accuracy.

The poster suggests that requiring models to answer honestly and stand behind their words may boost accuracy — as if models have their own "credibility score." A fun logit-intervening experiment hinting less hedging leads to more decisive, correct answers.

Related event: Penalizing "wait" and "maybe" boosts Qwen accuracy(4 posts)→

Original post →

More from Models

Models channel →