negative-trajectories

ReflectRL is a framework that learns from 'golden negative trajectories' (failed reasoning attempts by expert models) by reflecting on them, then transfers this reflective reasoning back to direct reasoning, improving LLM performance across benchmarks.