22:54
2026-09-27
startupfortune.com
large-language-models
A Reddit thread found that banning three words makes Qwen reasoning models sharper
A logit bias penalty of -2 applied to hedging tokens such as "wait," "maybe" and "perhaps" improved a Qwen3.5-4B model's accuracy on 50 MATH-500 questions while using fewer tokens, according to a r/Lo…