Large Language Models exhibit mode collapse, producing homogeneous outputs that fail to explore valid solution spaces. We present QD-LLM, a framework for parameter-efficient neuroevolution that evolves prompt embeddings, compact neural interfaces (~32K parameters) that steer generation in frozen LLMs (70B+ parameters), within a Quality-Diversity (QD) optimization framework. Our contributions: (1) evolved prompt embeddings via gradient-free optimization enabling behavioral steering without model fine-tuning; (2) hybrid behavior characterization combining semantic and explicit features with formal coverage bounds (Theorem 1) under validated near-independence (NMI $= 0.08 \pm 0.02$); (3) co-evolutionary variation operators including targeted behavioral mutation via finite-difference gradient estimation. On HumanEval (164 problems), MBPP, and creative writing benchmarks, QD-LLM achieves 46.4% higher coverage and 41.4% higher QD-Score than QDAIF ($p<0.001$, 30 runs, Vargha-Delaney $A=0.94$). We demonstrate downstream utility: diverse archives improve test generation (34% more edge cases) and fine-tuning data quality (8.3% accuracy gain). We validate across open-source LLMs (Llama-3-70B, Mistral-Large) with full embedding access, establishing prompt embedding evolution as an effective paradigm bridging neuroevolution and modern LLMs.

翻译：大语言模型存在模式坍塌问题，生成的同质化输出难以探索有效的解决方案空间。我们提出QD-LLM框架，采用参数高效的神经进化方法，在质量-多样性（QD）优化框架内进化提示嵌入（约32K参数的紧凑神经接口），以引导冻结的大语言模型（700亿+参数）生成。本文贡献包括：（1）通过无梯度优化实现提示嵌入进化，无需微调模型即可引导行为；（2）混合行为表征方法，结合语义特征与显式特征，并在验证近似独立性（归一化互信息NMI $= 0.08 \pm 0.02$）条件下建立形式化覆盖界（定理1）；（3）协同进化变异算子，通过有限差分梯度估计实现目标行为突变。在HumanEval（164个问题）、MBPP及创意写作基准测试中，QD-LLM相比QDAIF方法实现46.4%的覆盖率提升和41.4%的QD-Score提升（$p<0.001$，30次运行，Vargha-Delaney $A=0.94$）。我们展示了下游应用价值：多样性存档改善测试生成（增加34%边缘案例）和微调数据质量（准确率提升8.3%）。在具有完整嵌入访问权限的开源大语言模型（Llama-3-70B、Mistral-Large）上的验证表明，提示嵌入进化是连接神经进化与现代大语言模型的有效范式。