[4M5-GS-2f] Machine learning

4:30 PM - 4:45 PM JST(7:30 AM - 7:45 AM UTC)

[4M5-GS-2f-05] Joint Adversarial Prompting for Reflective Prompt Optimization

〇Kotaro Inoue¹ (1. SmartHR, Inc.)

Keywords:

Adversarial training,Hallucination Detection,Prompt Optimization

Prompting critically affects the performance of large language models (LLMs). In many practical applications, however, only a few labeled examples are available, making prompt optimization difficult under cold-start and cost constraints. In this paper, we propose Adversarial GEPA, an adversarial extension of GEPA, an reflection-based framework that improves prompts with a small number of iterations. Adversarial GEPA jointly optimizes the target prompt and adversarial generation prompts that synthesize challenging evaluation instances. After each update of the target prompt, adversarial samples are regenerated to progressively harden the evaluation set, yielding a co-evolutionary optimization process. Experiments on hallucination detection show that Adversarial GEPA can start from as few as four labeled examples and outperforms existing methods across multiple metrics, including AUROC and F1.