Distilling LLM Prior to Flow Model for Generalizable Agent’s Imagination in Object Goal Navigation | Read Paper on Bytez