Preprint
Scalable In-Context Reinforcement Learning with Recurrent Algorithm Distillation
Proposed Recurrent Algorithm Distillation (RAD), a dual-memory in-context RL agent that recursively compresses long interaction histories into a fixed-size latent memory. The policy autoregressively predicts actions conditioned on this memory plus a short working memory.
Across multiple environments, RAD matches or outperforms AD with up to 4× shorter context and up to 93% fewer inference FLOPs per action.