TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization
TextReg curbs prompt overfitting, gaining up to 16.5% out-of-distribution accuracy over REVOLVE.
TextReg addresses prompt distributional overfitting, in which LLM-rewritten prompts become longer, accumulate narrow rules, and generalize poorly beyond training data. The authors measure representational inefficiency as coupled capacity cost and scope narrowness, then apply Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Prompt Update. Across reasoning benchmarks, TextReg improves out-of-distribution accuracy by up to 11.8% over TextGrad and 16.5% over REVOLVE.
- Iterative prompt rewrites grow longer and overfit narrow sample-specific rules.
- TextReg penalizes capacity cost and scope narrowness with regularized textual gradients.
- Out-of-distribution gains reach +11.8% over TextGrad and +16.5% over REVOLVE.
From the original article
1 min read at huggingface.co
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.
By Lucheng Fu, Ye Yu, Yiyang Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.