• Diffusion-based framework refines system prompts via masked denoising in an iterative manner. • Conditions on interaction traces: user queries, model responses, and optional feedback. • Enables span-level prompt updates without gradient access or modifying downstream LLM. • Improves frozen target LLM performance on benchmarks like τ-bench, SST-2, SST-5. • Moderate diffusion step counts balance refinement quality and stability. • Demonstrates model-agnostic, scalable prompt optimization for enhancing LLM capabilities.

Article Summaries:

  • Prompt Optimization Via Diffusion Language Models Researchers have introduced a diffusion‑based framework that refines system prompts for large language models (LLMs) without requiring gradient access or altering the target model. By applying masked denoising steps to Diffusion Language Models (DLMs) and conditioning on interaction traces-user queries, model responses, and optional feedback-the method iteratively updates prompt spans. Experiments on benchmarks such as τ‑bench, SST‑2, and SST‑5 show consistent performance gains for frozen LLMs like GPT‑4o‑mini. The study finds that a moderate number of diffusion steps balances refinement quality and stability, positioning diffusion‑driven prompt optimization as a general, scalable, and model‑agnostic enhancement technique.

Sources: