TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization

Can prompts overfit just like machine learning models do?

Prompt optimization has emerged as a powerful paradigm for improving LLM performance. However, we observed a recurring phenomenon: as optimization progresses, prompts often become longer, accumulate increasingly specific instructions, and generalize worse beyond the training distribution.

In this work, we study this failure mode as prompt distributional overfitting.

Rather than viewing prompts as arbitrary text, we treat them as structured representations of task knowledge and ask a fundamental question:

What makes an optimized prompt generalize?

To answer this question, we introduce representational inefficiency, a perspective that attributes prompt overfitting to the coupled growth of:

🔹 Capacity Cost — prompts becoming unnecessarily long

🔹 Scope Narrowness — rules becoming increasingly specific and less reusable

Building on this insight, we propose TextReg, a regularization framework for text-space prompt optimization that combines:

✅ Dual-Evidence Gradient Purification
✅ Semantic Edit Regularization
✅ Regularization-Guided Prompt Update

Across multiple reasoning benchmarks and model families, TextReg consistently improves out-of-distribution generalization, achieving gains of up to +11.8% over TextGrad and +16.5% over REVOLVE.

More broadly, we hope this work contributes to a better understanding of how prompts evolve during optimization and how regularization principles can be extended from parameter space to text space.