Some research does not sit inside a single discipline. The work below comes from Toronto-area laboratories where an engineering method and a very different field had to be reconciled before the result could exist at all. 

Four agents, one psychiatric interview 

Structured psychiatric interviews improve diagnostic reliability, and they are expensive to administer. A Toronto team led from the AI for Mental Health Program at St. Michael’s Hospital has built a multi-agent large language model framework that administers the protocol itself, published in Translational Psychiatry on 31 July 2026. 

Four agents divide the work: a Questioner administers the interview modules, an Evaluator interprets free-text answers, a Navigator enforces bounded clarification loops, and a Diagnoser applies encoded decision rules to produce a screening outcome. Personally identifiable and protected health information is redacted before any model processing. Across 1,350 simulated interviews spanning 15 neuropsychiatric screening modules, module-level decisions agreed with rule-based expected outcomes 87.8% of the time, at 88.9% sensitivity and 86.7% specificity. On a separate 2,000-sentence red-flag benchmark the classifier reached 93.9% accuracy. The multi-agent design scored an F2 of 88.5% where a single-pass model managed 49.5%. 

The author list runs from the Department of Electrical, Computer, and Biomedical Engineering at Toronto Metropolitan University through that university’s Department of Psychology and York University’s Department of Mathematics and Statistics to clinical psychiatry at the Centre for Addiction and Mental Health and Unity Health Toronto. The engineering contribution is not the language model. It is the orchestration, the bounded loops, the escalation path, and the auditability that make a generative system admissible in a clinical protocol at all. 

doi:10.1038/s41398-026-04335-5 

Testing machine explanations on actual brains 

Explainability methods for artificial neural networks are cheap to compute, frequently disagree with one another, and are almost never checked against a biological system. Researchers at York University’s Centre for Vision Research, with the Department of Biomedical Engineering at the University of Toronto, published a framework in Communications Psychology on 13 July 2026 that closes that loop. 

The method — Masked Attribution-based Probing of Strategies, or MAPS — converts an explanation map into a minimal explanation-masked image and asks whether that image still produces the same recognition behaviour as the original. If the explanation captured what the system actually uses, it should. Tested against 56 human participants and two macaques, MAPS identified which explanation methods align with biological vision at a validity comparable to the established Bubbles psychophysics technique, without the exhaustive data collection Bubbles requires. 

This is a machine-learning evaluation problem answered with a primate-neuroscience experiment. Neither half works alone. 

doi:10.1038/s44271-026-00502-y 

Authors Bio 

Alexander Chernov is an Associate Principal Data Engineer in Data Science & Modelling, BioPharmaceuticals R&D, at AstraZeneca in Toronto, where he architects governed, AI-enabled data platforms serving biomedical researchers worldwide under regulated (GxP) controls. His research addresses secure AI infrastructure — semantic caching, policy-aware agent architectures, and observability for autonomous systems