LANTERN: LLM-Augmented Neurosymbolic Transfer Learning Framework

A novel reinforcement learning framework named LANTERN (LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks) has been introduced to enhance transfer learning among various source tasks. This framework overcomes the shortcomings of current neurosymbolic approaches by utilizing large language models to create deterministic finite automata from task descriptions in natural language. This allows for the aggregation of multiple source policies based on semantic embedding and cross-task similarity. Additionally, it features adaptive teacher-student gating informed by temporal-difference error and semantic uncertainty. In tests conducted in resource management, navigation, and control areas, LANTERN demonstrated a 40-60% increase in sample efficiency compared to baseline methods. The research can be found on arXiv with the identifier 2605.05478.