Reflective Prompt Tuning through Language Model Function-Calling

Researchers introduced Reflective Prompt Tuning, a framework that uses an LLM to diagnose failures on an evaluation set and iteratively revise prompts.

The method combines structured diagnostic feedback with prior diagnostic memory to refine prompts more systematically.

It improved performance by up to 12.9 points across three reasoning tasks and also boosted confidence calibration.

The approach reduces manual prompt-tuning effort while making large language models more reliable on complex reasoning tasks.