Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization

TL;DR

This paper introduces VISTA, a multi-agent framework for transparent and effective prompt optimization in large language models, overcoming black-box limitations and improving accuracy on benchmark datasets.

Contribution

VISTA decouples hypothesis generation from prompt rewriting, enabling interpretability and parallel verification, with a novel explore-exploit strategy to avoid local optima.

Findings

VISTA recovers accuracy to 87.57% on defective seed GSM8K.

VISTA outperforms baselines across GSM8K and AIME2025.

Identifies limitations of existing reflective APO methods.

Abstract

Automatic prompt optimization (APO) has emerged as a powerful paradigm for improving LLM performance without manual prompt engineering. Reflective APO methods such as GEPA iteratively refine prompts by diagnosing failure cases, but the optimization process remains black-box and label-free, leading to uninterpretable trajectories and systematic failure. We identify and empirically demonstrate four limitations: on GSM8K with a defective seed, GEPA degrades accuracy from 23.81% to 13.50%. We propose VISTA, a multi-agent APO framework that decouples hypothesis generation from prompt rewriting, enabling semantically labeled hypotheses, parallel minibatch verification, and interpretable optimization trace. A two-layer explore-exploit mechanism combining random restart and epsilon-greedy sampling further escapes local optima. VISTA recovers accuracy to 87.57% on the same defective seed and consistently outperforms baselines across all conditions on GSM8K and AIME2025.