ReflectRL uses failed expert trajectories for reasoning improvement
August 3, 2026
ReflectRL improves LLM reasoning by treating failed expert trajectories as Golden Negative Trajectories rather than discarding them. The method leverages a Reflection Advantage, where models learn more effectively by reflecting on flawed reasoning paths than by attempting to solve hard problems from scratch.
HOW THIS AFFECTS YOU
●
builderYou can improve model reasoning by implementing reflection-based learning on difficult tasks.
●
researcherYou can utilize failed training samples to enhance post-training reasoning capabilities.