espl-evolutionary-system-prompt

What problem does it solve?

Dual optimization of system prompts and model weights using RL and evolutionary methods to enable agent self-improvement.

Core Features & Use Cases

• Dual optimization loop combining RL updates for weights with evolutionary prompt mutations.
• Genetic operators for prompts: mutation and crossover via LLM reflections.
• Clear separation of declarative knowledge (prompts) and procedural knowledge (weights) to support scalable agent improvement.
• Use cases include self-improving agents, multi-task optimization, and long-horizon learning tasks.

Quick Start

Initialize a population of prompts, run parallel trajectories to evaluate them, perform RL weight updates, apply prompt evolution, and iterate until convergence.