Hybrid SFT-RL / Prefix-RFT (UFT / LUFFY / Prefix-Guided RLVR)

When a hard problem has $0\%$ pass rate under current policy sampling ($\text{pass}@G = 0$), pure RLVR receives zero positive signal. Hybrid SFT-RL (LUFFY / UFT / Prefix-RFT) injects off-policy expert prefixes $y_{1:k}^\star$ (from human/teacher traces) or mixes off-policy expert trajectories into the GRPO group using regularized importance sampling, smoothly annealing prefix length $k \to 0$ as the policy learns to solve the task on its own.