sparse-reward

Proposes ULPS, a framework integrating a calibrated LLM into RL training with uncertainty-modulated guidance and A*-based symbolic trajectories, achieving improved success rate and sample efficiency on MiniGrid-UnlockPickup.