LANTERN: LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks

Researchers introduce LANTERN, a framework that uses large language models to automatically generate task descriptions and intelligently aggregate knowledge from multiple source tasks for reinforcement learning. The system achieves 40-60% improvements in sample efficiency by adaptively weighting source policies based on task similarity and managing teacher-student knowledge transfer through uncertainty-aware gating.

LANTERN addresses a fundamental challenge in reinforcement learning: efficiently transferring knowledge across related tasks without manual engineering. Traditional neurosymbolic transfer approaches require humans to hand-code task automata and typically handle only single source tasks, creating bottlenecks in scalability and practical deployment. This research demonstrates how large language models can bridge the gap between natural language task descriptions and formal computational structures, eliminating tedious manual specification while enabling multi-source knowledge aggregation.

The framework's innovation lies in three integrated mechanisms working in concert. Natural language task descriptions become deterministic finite automata through LLM interpretation, solving the manual specification problem. Rather than treating all source tasks equally, LANTERN weights multiple policies by semantic similarity, allowing the system to recognize when sources are irrelevant or misaligned. The adaptive gating mechanism represents particularly sophisticated design—using temporal-difference error and semantic uncertainty to dynamically adjust how much to trust source policies versus learn independently, preventing negative transfer from poorly matched sources.

The experimental results across resource management, navigation, and control domains show consistent 40-60% sample efficiency gains, suggesting the approach generalizes beyond narrow domains. Robustness to misaligned sources distinguishes LANTERN from brittle transfer methods that fail catastrophically when source tasks diverge from target tasks.

For AI research communities, this work demonstrates the practical value of combining symbolic reasoning with neural learning and LLM capabilities. The approach reduces friction in applying transfer learning to new domains by automating specification while maintaining interpretability. Future developments likely involve extending this framework to more complex task hierarchies and real-world robotics applications where manual tuning remains prohibitively expensive.

• →LANTERN uses LLMs to automatically convert natural language task descriptions into formal automata, eliminating manual specification overhead.
• →Multi-source policy aggregation weighted by semantic similarity improves robustness by reducing negative transfer from misaligned source tasks.
• →Adaptive teacher-student gating based on uncertainty metrics achieves 40-60% sample efficiency gains across multiple domains.
• →The framework bridges neurosymbolic reasoning with modern deep learning, combining interpretability with neural function approximation.
• →Demonstrated robustness to poorly aligned sources suggests practical applicability to real-world transfer learning scenarios.