confidence-rl-finetuning

【数据与研究】Improve language model reasoning using only model confidence as reward signals, eliminating need for labels or preference models while achieving substantial gains with minimal data.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准