I work on generative AI systems that are supposed to produce something new — and on the mechanisms by which they stop doing so. The three directions below are one question seen from three sides.
Current directions
Systems that generate and test hypotheses on their own, rather than executing a plan a human wrote out in advance. Search over ideas, self-directed experimentation, and the evaluation problem underneath both: deciding whether an agent's output is a genuine finding or a fluent restatement of its training data. My own attempts at this have been in biomarker discovery — searching for candidate markers in retinal imaging rather than validating a pre-specified list.
Agents that rewrite their own scaffolding — prompts, tools, the criteria they are judged by — and improve across iterations. The interesting failure is not that such a loop stalls but that it can appear to improve while drifting: when the evaluator co-evolves with the agent, the reward can move faster than the capability. What keeps a self-improvement loop anchored is, to me, the open question.
What happens to a generative model's diversity once it trains on machine-generated output. This is the failure mode underneath the creativity question, and it is not gradual in the way one might hope: the rare cases go first. Across more than 800,000 synthetic data points spanning clinical text, vision–language reporting and medical image synthesis, we found models converging on generic phenotypes irrespective of architecture, with rare but critical findings disappearing from the output entirely.
Why medicine
Clinical settings are where I test these questions, not the subject of them. Two reasons. A lost rare case is not an abstract diversity metric there — it is a missed pneumothorax. And medical records are already being populated with AI-generated content, which makes the contamination loop something to measure now rather than forecast.
Other threads
A survey of the safety challenges specific to large language models in medicine, and cross-lingual work showing that diagnostic safety is language-dependent — performance and failure modes shift with the language a system is queried in, which matters for anywhere that is not monolingual English.
At the Department of Radiology, Duke University Medical Center. Multi-label classification of bone type and MRI series type (axial / sagittal / coronal), annotation and 3D segmentation of bone MRI, and CT-based body composition analysis for disease-risk prediction. The SuperMask work came out of this period: reconstructing high-resolution object masks from multi-view, unaligned low-resolution MRIs.
Stochastic channel-based federated learning with neural network pruning, aimed at training on medical data without centralising it; and FaceGuard, a proactive approach to deepfake detection.
OpenHI2, an open-source platform for histopathological image annotation and analysis.
If any of this overlaps with what you are working on, I am glad to hear from you — hongyu.h@nus.edu.sg.