Study: Weaker LLMs Rewriting Prompts for Stronger Models Boosts Zero-Shot Performance

max_paperclips · x · 2026-08-05

The paper introduces PRomPTed, an approach that leverages LLMs in the loop to rewrite prompts for individual task instances, optimizing zero-shot performance.

Key Findings: Evaluated across 13 datasets and 10 task types, the method outperforms both naive zero-shot approaches and output refinement baselines.
Counterintuitive Insight: Using the weaker GPT-3.5 to rewrite prompts for the stronger GPT-4 matches or occasionally exceeds the efficacy of using GPT-4 itself as the rewriter.

This highlights a valuable strategy for reducing API costs and enabling weaker models to supervise stronger ones.

Original post →

Study: Weaker LLMs Rewriting Prompts for Stronger Models Boosts Zero-Shot Performance

More from coding & agent