LUFFY: Learning to Reason under Off-Policy Guidance

An RL-for-reasoning method from Yue Zhang's WestlakeNLP that mixes off-policy expert traces into RLVR, letting a model learn reasoning behaviors beyond what its own base policy can reach. A heavily-cited (~240) foundational RL-reasoning technique. (Yue Zhang also co-authored the SuperGPQA knowledge/reasoning benchmark, tracked under ByteDance Seed.)