What does it take to bridge the gap between today’s code-generating systems (e.g., Copilot, Cursor) and a real human programmer? I build interactive code-generation (broadly, agentic) systems that communicate and collaborate with end users to solve tasks. These systems are intuitive to instruct while reliably interpreting and executing user intent. To build such systems, we draw insights from human communications, especially the cognitive/linguistic study on how humans achieve common-ground understanding through dialogue. Operatinally, I (a) curate human-human agentic datasets and benchmarks that remain challenging for today’s best agentic systems, (b) develop novel Bayesian models of human communication from (a), and build SOTA interactive code-generation systems from (b) .
I started as an asst prof at NTU in June 2025. Previously, I was a principal research scientist at Autodesk AI Lab, where I work on code-generation for human-machine collaboration in CAD, and industry scale instruction-following dataset annotation. My domain expertise include code generation, embodied instruction following agents, and collaborative design. My target communities span broadly across Cogsci, NLP, PL, Graphics, and HCI. I received my PhD under Armando Solar-Lezama and Leslie Kaelbling at MIT.
Representative Works
- Program Synthesis with Pragmatic Communication (NeuRIPS 2020)
Yewen Pu, Kevin Ellis, Marta Kryven, Josh Tenenbaum, Armando Solar-Lezama - Communicating Natural Programs to Humans and Machines (NeuRIPS 2022)
Samuel Acquaviva *, Yewen Pu *, Marta Kryven, Theodoros Sechopoulos, Catherine Wong, Gabrielle E Ecanow, Maxwell Nye, Michael Henry Tessler, Joshua B. Tenenbaum - mrCAD: Multimodal Refinement of Computer-aided Designs (arXiv 2025)
William P. McCarthy, Saujas Vaduguru, Karl D. D. Willis, Justin Matejka, Judith E. Fan, Daniel Fried, and Yewen Pu
Additional Information
Shoot me a twitter message or email any time if you want to chat !