Meta-Morphic Evaluation Networks: Ground-Truth-Free Robustness Testing via Logical Prompt Pairings

Shen et al. (2025) introduce MetaLogic for text-to-image models, using logically equivalent prompts to uncover semantic misalignment. This research extends and generalizes this idea to language models and multimodal systems, constructing “meta-morphic networks” where prompts are linked by logical, paraphrastic, or semantic relations. By traversing these networks and evaluating model consistency across chains of equivalent prompts, we can identify subtle robustness failures—such as drift, compounding errors, or overlooked biases—without requiring ground-truth answers. This method not only scales metamorphic testing, but also helps surface model “blind spots” that evade conventional evaluation. Such a framework could become a new standard for model auditing and benchmarking.

References:

1. MetaLogic: Robustness Evaluation of Text-to-Image Models via Logically Equivalent Prompts. Yifan Shen, Yangyang Shu, Hye-young Paik, Yulei Sui (2025).