RLPE: Teaching LLMs to Prompt with Reinforcement Learning

Recent research introduces RLPE and MemComposer to enhance LLM capabilities through Reinforcement Learning. RLPE focuses on Reinforcement Learned Prompt Engineering, training models to create effective prompts by observing task model behavior, avoiding the overhead of inference-time systems. Simultaneously, MemComposer improves LLM agents by optimizing memory assembly. Using the ExPO framework based on GRPO, MemComposer determines the selection, quantity, and ordering of retrieved memories to ensure they are useful for downstream performance and agent interactions.