Machine learning systems are usually assumed to face the same information at training time and at inference time, but many real-world scenarios break that symmetry. Additional or higher-quality information may be available only during training, and no longer at inference time.
The survey, authored by Veronica Pignedoli, Matteo Moro, Nicoletta Noceti and Francesca Odone, reviews 107 papers published between 2009 and 2025. The authors describe the LUPI literature as highly fragmented despite years of research interest, and their stated goal is a comprehensive and structured overview that unifies the paradigm with its related formulations.
The core contribution is a novel unified taxonomy built on two dimensions. The second concerns the methodological strategies used to integrate privileged information into the learning process.
Beyond the taxonomy, the survey provides an extensive overview of the application domains where LUPI has been applied and catalogues the particular types of privileged modalities involved.
The paper closes with a discussion of key insights and open challenges, centred on a question the authors frame directly: how and when can privileged information be effectively leveraged in modern machine learning systems? The article was published open access on 28 October 2026 as a peer-reviewed, accepted manuscript carrying a permanent DOI, with a note that this early version remains subject to further edits before the final Version of Record.
Subject of Research: Learning Under Privileged Information in machine learning
Article Title: Learning under privileged information: a survey
Article References: Pignedoli, V., Moro, M., Noceti, N., & Odone, F. (2026). Learning under privileged information: a survey. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11746-9
Image Credits: AI Generated
DOI: 10.1007/s10462-026-11746-9
Keywords: machine learning, privileged information, LUPI, survey, taxonomy, deep learning, training-inference asymmetry, Artificial Intelligence Review, Learning, under, privileged, information
News Source: Blake Davidson. (October 28, 2026). Survey maps how machine learning exploits privileged information. Scienmag.
