confidence-rl-finetuning

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.

Independent GitHub snapshot · Automated classification · Original documentation governs