Active Diffusion Model Debugging: Generative Models that Suggest their Own Dataset or Architecture Improvements

While Liu et al. (2025) highlight the reinforcing cycle between detection and generation in anomaly detection, the loop typically ends with detection or sample generation. Here, I propose a generative framework where, upon detecting anomalous behavior or generation failures (see also “unexpected behavior” heuristics), the diffusion model suggests concrete improvements—such as creating synthetic samples to augment underrepresented regions of the dataset, or even proposing architecture changes (like adding attention heads or modifying loss functions). This “active debugging” could leverage techniques from explainable AI (Nabeel et al., 2024), and would represent a shift from passive to proactive model improvement—a meta-generative approach not found in current literature.

References:

1. Fault Detection in Mobile Networks Using Diffusion Models. Mohamad Nabeel, Doumitrou Daniil Nimara, Tahar Zanouda (2024). 2024 IEEE International Conference on Communications Workshops (ICC Workshops).
2. Anomaly Detection and Generation with Diffusion Models: A Survey. Yang Liu, Jing Liu, Chengfang Li, Rui Xi, Wenchao Li, Liang Cao, Jin Wang, Laurence T. Yang, Junsong Yuan, Wei Zhou (2025). arXiv.org.