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

TL;DR

This paper introduces TextReg, a regularization framework for prompt optimization in large language models that mitigates overfitting and improves out-of-distribution generalization by controlling representation growth.

Contribution

It formalizes prompt overfitting as representational inefficiency and proposes a novel regularization method combining multiple gradient-based techniques.

Findings

TextReg improves OOD generalization with up to +16.5% accuracy gains.

It effectively controls prompt overfitting during optimization.

The framework enhances reasoning benchmark performance.

Abstract

Large language models (LLMs) are highly sensitive to the prompts used to specify task objectives and behavioral constraints. Many recent prompt optimization methods iteratively rewrite prompts using LLM-generated feedback, but the resulting prompts often become longer, accumulate narrow sample-specific rules, and generalize poorly beyond the training distribution. We study this failure mode as prompt distributional overfitting and argue that it reflects a lack of representation control in discrete text-space optimization. We formalize this view through representational inefficiency, a dual-factor measure that decomposes prompt inefficiency into capacity cost and scope narrowness, attributing distributional prompt overfitting to their coupled growth during optimization. We propose TextReg, a regularization framework that realizes a soft-penalty objective through regularized textual gradients, combining Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Prompt Update. Across multiple reasoning benchmarks, TextReg substantially improves out-of-distribution (OOD) generalization, with accuracy gains of up to +11.8% over TextGrad and +16.5% over REVOLVE.