Espl Evolutionary System Prompt Learning

---\nname: espl-evolutionary-system-prompt-learning\ndescription: 'Evolutionary System Prompt Learning (E-SPL) jointly improves model context (system prompt) and model weights through parallel sampling and LLM self-reflection driven evolution. Based on arXiv:2602.14697.'\n---\n\n# Evolutionary System Prompt Learning (E-SPL)\n\n**arXiv**: 2602.14697 | **Utility**: 0.9\n\n## Overview\nEvolutionary System Prompt Learning (E-SPL) jointly improves model context (system prompt) and model weights through parallel sampling of trajectories under different system prompts and LLM self-reflection driven evolution of the prompts. The system prompt evolves through mutation and crossover operations, encouraging declarative knowledge encoding in prompts and procedural knowledge encoding in weights.\n\n## Key Contributions\n\n1. Joint optimization of system prompt and model weights through evolutionary algorithms\n2. Parallel sampling of multiple system prompt variants during RL training\n3. LLM self-reflection drives the mutation and crossover of system prompts\n4. Encourages declarative knowledge encoding in prompts and procedural knowledge in weights\n5. Demonstrated improvement: RL success rate increased from 38.8% to 45.1%\n\n## Implementation Notes\n\n- **Keywords**: self-evolution, system-prompt, RL\n- **Categories**: cs.LG, cs.AI, cs.CL\n- **Published**: 2026-02\n\n## Activation Criteria\n\nUse this skill when working on tasks involving: self-evolution, system-prompt, reinforcement-learning, prompt-optimization.\n\n## Implementation Approach\n\n1. **Dual Optimization**: Simultaneously optimize both the system prompt (context) and model weights\n2. **Population-Based Search**: Maintain a population of system prompt variants\n3. **Parallel Rollouts**: Sample trajectories for each prompt variant in parallel during RL\n4. **Self-Reflection Evolution**: Use LLM to critique and mutate prompts based on performance\n5. **Knowledge Separation**: Encourage factual/declarative knowledge in prompts, procedural skills in weights\n\n## Application Notes\n\n- Particularly effective for reasoning and planning tasks\n- Can be applied to improve agent system prompts like SOUL.md and AGENTS.md\n- Complements weight-based fine-tuning with prompt-based adaptation\n- Enables continual improvement through evolutionary pressure on the instruction space\n\n## References\n\n- arXiv:2602.14697 - Evolutionary System Prompt Learning (E-SPL)\n\n---\n*Skill created from arXiv paper tracking system. Utility: 0.9*\n