QD-LLM: Parameter-Efficient Neuroevolution for Diverse LLM Generation

Researchers have introduced QD-LLM, a novel framework addressing mode collapse in Large Language Models (LLMs) by employing parameter-efficient neuroevolution. Instead of fine-tuning massive frozen models, QD-LLM evolves compact prompt embeddings (~32K parameters) to steer generation within a Quality-Diversity (QD) optimization framework. The method utilizes gradient-free optimization and hybrid behavior characterization to ensure diverse and high-quality outputs. Validated on benchmarks like HumanEval and MBPP using Llama-3-70B and Mistral-Large, QD-LLM demonstrated significant improvements, achieving 46.4% higher coverage and a 41.4% higher QD-Score compared to existing methods. The study highlights downstream benefits, including a 34% increase in edge case detection for test generation and an 8.3% accuracy gain in fine-tuning data quality. This research establishes prompt embedding evolution as an effective paradigm for bridging neuroevolution with modern LLMs, offering a scalable solution for enhancing output diversity without the computational cost of full model retraining.