LUFFY
A reinforcement learning training method for large language models that incorporates external guidance, such as expert trajectories, to accelerate or improve multi-step reasoning training by injecting external knowledge into the optimization process.
Timeline (last 90 days)
GA-GRPO abstracts external guidance as a stochastic guidance operator G; by choosing different G, vanilla GRPO, LUFFY, ExPO, PAPO, and TAPO all become special cases of this framework.
GA-GRPO with optimal weights matches or surpasses TAPO, LUFFY, ExPO, and vanilla GRPO in performance, while requiring 31% fewer GPU hours.
Methods such as LUFFY, ExPO, PAPO, and TAPO enhance large model reasoning by introducing external guidance, including expert trajectories, self-explanations, or retrieved reasoning patterns.
All Facts (3)
GA-GRPO abstracts external guidance as a stochastic guidance operator G; by choosing different G, vanilla GRPO, LUFFY, ExPO, PAPO, and TAPO all become special cases of this framework.
50%UnverifiedGA-GRPO with optimal weights matches or surpasses TAPO, LUFFY, ExPO, and vanilla GRPO in performance, while requiring 31% fewer GPU hours.
50%UnverifiedMethods such as LUFFY, ExPO, PAPO, and TAPO enhance large model reasoning by introducing external guidance, including expert trajectories, self-explanations, or retrieved reasoning patterns.
50%