GenAI Research

Undergraduate Honours Thesis

A central goal in controllable generative modeling is to steer a target concept through latent interventions without inadvertently changing other concepts. A natural approach is to use probes, but probe accuracy is not a reliable indicator of steerability. We introduced VE1*, a post-hoc steerability metric grounded in decoder Jacobian geometry, outperforming linear probes and sparsity baselines. We tested across various generative model types like VAEs, GANs, and flow matching. Thank you to my supervisors Dr. Mo Chen (Associate Prof, Canada CIFAR AI Chair, SFU) and Dr. Carl Allen (Laplace Chair in ML, ENS Paris). The code is private, but you can still see the condensed paper (full paper is hidden since it's not on arXiv yet). Project website →