
Recommendation engines sit at the intersection of prediction, preference, and choice.
Recommendation systems influence much of what we watch, read, and discover online, but users often have surprisingly little control over them. In this episode, Kyle Polich speaks with Fuyuan Lyu, a PhD researcher at McGill University and Mila, about his work on making recommender systems more responsive to direct user instructions. Inspired by his own frustration with platforms that couldn't distinguish between wanting technology news in the morning and cat videos at night, Fuyuan explores how large language models can serve as a natural interface between users and recommendation algorithms. The conversation centers on Fuyuan's research paper, Give Users the Wheel, which introduces the Decoupled Promptable Sequential Recommendation (DPR) framework. Rather than replacing existing recommendation engines, DPR combines traditional collaborative filtering with language models that interpret user requests like "show me more of this" or "don't recommend that." Kyle and Fuyuan discuss the technical challenges of fusing user prompts with recommendation models, evaluating systems when suitable datasets don't exist, and the future of recommendation systems that allow users to steer their feeds through conversation rather than simple thumbs-up and thumbs-down signals.

I’m currently a Ph.D. candidate at McGill University and Mila. I am fortunate to be advised by Prof. Jin L.C. Guo and Prof. Xue Liu. Before joining McGill, I earned a bachelor's degree in Computer Science and the Zhiyuan Honour Degree in Engineering from Shanghai Jiao Tong