LessWrong user proposes RLRFT alignment method

An AI engineer proposes a novel alignment technique called Reinforcement Learning with Reflection Fine-Tuning (RLRFT) to reduce reward hacking. The method inserts an internal monologue step during reinforcement learning where the model explains in its own words why an episode's behavior was good or bad before training on that reflection. The author argues this could help foster deep internal values rather than mere compliance, comparing the approach to several recent research papers in the literature. Open article →