Vincent Sunn Chen

Chen did graduate work at the Stanford AI Lab with Chris Ré, studying multi-task learning, weak supervision and long-tail evaluation. Earlier, he worked on Tesla Autopilot’s data engine and co-directed TreeHacks. His current focus includes benchmarks, scalable expert feedback and post-training feedback loops. He was a teaching assistant for Stanford’s CS231N in spring 2018.[1][2]

Give critical slices additional model capacity

His coauthored Slice-based Learning paper starts from failures hidden by an aggregate score. Developers write noisy slicing functions identifying important subsets, such as rare road users or particular language patterns. The model learns slice-specific representations using a shared backbone, then combines them with attention. This allocates capacity to the critical cases without training a separate full model for every subset. The method studies both slice performance and the overall task.[3]

Create relationship labels from sparse supervision

In Scene Graph Prediction with Limited Labels, Chen and his coauthors use a few labeled examples to construct noisy heuristics for visual relationships. A generative model combines those signals into probabilistic labels for previously unlabeled images. Those labels can then train a scene-graph model. The work examines where this form of weak supervision helps relative to transfer learning, connecting the structure of a relationship with the method’s effectiveness.[4]

For Benchtalks, Chen hosted John Yang in a conversation about the path from SWE-bench and InterCode to ProgramBench. The original announcement highlights artifact-level evaluation, benchmark reception and models’ preference for Python as discussion topics. This interviewing work extends his interest in how an evaluation’s design shapes the behavior it reveals.[5]