PromptSplit: Revealing Prompt-Level Disagreement in Generative Models

TL;DR

PromptSplit is a kernel-based framework that detects and analyzes prompt-dependent disagreements among generative models, providing interpretable insights into their behavioral differences across various tasks.

Contribution

It introduces a scalable, kernel-based method with theoretical guarantees for identifying prompt-driven model disagreements in generative AI.

Findings

Accurately detects ground-truth behavioral differences

Isolates prompts responsible for model disagreement

Provides an interpretable analysis of generative model behavior

Abstract

Prompt-guided generative AI models have rapidly expanded across vision and language domains, producing realistic and diverse outputs from textual inputs. The growing variety of such models, trained with different data and architectures, calls for principled methods to identify which types of prompts lead to distinct model behaviors. In this work, we propose PromptSplit, a kernel-based framework for detecting and analyzing prompt-dependent disagreement between generative models. For each compared model pair, PromptSplit constructs a joint prompt--output representation by forming tensor-product embeddings of the prompt and image (or text) features, and then computes the corresponding kernel covariance matrix. We utilize the eigenspace of the weighted difference between these matrices to identify the main directions of behavioral difference across prompts. To ensure scalability, we employ a random-projection approximation that reduces computational complexity to $O(nr^2 + r^3)$ for projection dimension $r$. We further provide a theoretical analysis showing that this approximation yields an eigenstructure estimate whose expected deviation from the full-dimensional result is bounded by $O(1/r^2)$. Experiments across text-to-image, text-to-text, and image-captioning settings demonstrate that PromptSplit accurately detects ground-truth behavioral differences and isolates the prompts responsible, offering an interpretable tool for detecting where generative models disagree.