RELIC Enables Private Skill Learning in Multi-Agent Systems

Multi-agent planning becomes significantly more complex when individual agents must develop and refine specialized decision-making skills while keeping their internal implementations confidential. This scenario is common when agents are developed independently, possess diverse interfaces and capabilities, yet need to coordinate effectively without direct policy sharing. Traditional approaches often assume centralized optimization or shared skill representations, which are unsuitable for privacy-constrained cooperation. The RELIC framework addresses this by enabling agents to learn interpretable and composable skills through "revealed principles." Each agent privately refines its programmatic skill using LLM-guided search. A trusted orchestrator then evaluates proposed updates based solely on the team's overall performance. Instead of broadcasting executable code, successful behaviors are abstracted into portable principles that other agents can instantiate within their own unique interfaces and combine with local strategies. This innovative approach separates coordination from implementation sharing, facilitating cross-agent transfer even with heterogeneous skill signatures, thus introducing a new paradigm for privacy-preserving skill learning and coordination in multi-agent systems.