TextReg: Regularized Text-Space Optimization for LLM Prompts

A new arXiv paper (2605.21318) introduces TextReg, a regularization framework to mitigate prompt distributional overfitting in large language models. The authors identify that prompt optimization methods often produce lengthy, narrow prompts that generalize poorly. They formalize this as representational inefficiency, measured by capacity cost and scope narrowness. TextReg applies a soft-penalty objective via regularized textual gradients, combining Dual-Evidence Gradient Purification to improve prompt robustness.