This research introduces a novel approach using reinforcement learning to automatically refine user queries, enhancing LLM performance while preventing security vulnerabilities.

• Creates more effective queries from vague or brief user prompts
• Significantly improves response quality and helpfulness
• Builds robust defenses against jailbreak attempts and harmful prompts
• Maintains model alignment with human values through guided refinement

For security teams, this research offers a crucial advancement in preventing adversarial manipulation of LLMs without sacrificing performance or user experience. The approach addresses the fundamental tension between capability and safety in deployed AI systems.

Enhancing the Capability and Robustness of Large Language Models through Reinforcement Learning-Driven Query Refinement