e-spl-evolutionary-system-prompt-learning

name	e-spl-evolutionary-system-prompt-learning
description	E-SPL (Evolutionary System Prompt Learning): Evolutionary system prompt learning for self-evolving LLMs. Uses reinforcement learning to optimize system prompts, enabling LLMs to improve themselves through interaction without external supervision. Activation: self-evolution, system-prompt, RL.
category	nlp-llm
metadata	{"arxiv_id":"2602.14697","published":"2026-02-24","authors":"Unknown","tags":["self-evolution","system-prompt","rl"]}

E-SPL (Evolutionary System Prompt Learning)

Overview

E-SPL introduces an evolutionary approach to system prompt learning where large language models can improve themselves through reinforcement learning without external supervision. The method optimizes system prompts that govern model behavior, enabling continuous self-improvement.

Core Methodology

Evolutionary Prompt Optimization

• System Prompt as Policy: Treat system prompts as policies in a reinforcement learning framework
• Fitness Function: Model performance on downstream tasks serves as reward signal
• Mutation & Crossover: Generate prompt variants through linguistic transformations
• Selection: Retain prompts that improve model performance

Self-Evolution Loop

1. Prompt Generation: Create diverse system prompt candidates
2. Model Adaptation: Adapt LLM behavior using each candidate prompt
3. Performance Evaluation: Evaluate adapted models on target tasks
4. Selection & Reproduction: Select high-performing prompts for next generation
5. Iteration: Repeat until convergence or budget exhaustion

Key Innovations

• No External Supervision: Uses model's own performance as reward signal
• Prompt Space Exploration: Effectively searches discrete prompt space
• Behavioral Adaptation: Changes model behavior through prompt modification
• Continuous Improvement: Enables ongoing self-improvement cycles

Activation Keywords

• self-evolution
• system-prompt
• rl
• evolutionary-prompt-learning
• self-improving-llm
• prompt-optimization

Implementation Notes

Requires access to model logits or probability distributions for reward computation. Works best with open-weight models where likelihood computation is feasible.

References

• arXiv:2602.14697 - E-SPL (Evolutionary System Prompt Learning)