espl-evolutionary-system-prompt

Co-evolving system prompts and model weights with RL improves agent performance by running a dual optimization loop that applies evolutionary prompt mutations alongside reinforcement learning updates. This enables continual self-improvement for multi-task and long-horizon decision tasks.

The best way to optimize system prompts and model weights simultaneously is to initialize a prompt population, evaluate them via parallel trajectories, perform RL weight updates, and apply LLM reflection-based genetic mutations for prompt evolution until convergence.

Yes, you need an RL framework like PPO or GRPO for evolutionary prompt optimization. You also require a prompt management mechanism, a knowledge graph to store evolution history, and an LLM to perform reflection-based genetic mutations like crossover.

Yes, you can use evolutionary prompt mutation for long-horizon learning tasks. By maintaining a clear separation of declarative prompt knowledge and procedural weight knowledge, the system supports scalable agent improvement across complex, long-horizon environments.

Separating declarative prompt knowledge from procedural weight knowledge in RL agents supports scalable agent improvement. It allows genetic operators to independently mutate prompts via LLM reflections while RL updates handle procedural weights, enabling robust multi-task optimization.