Date: October 1, 2026
Time: 11:45 a.m. - 12:45 p.m.
Location: G01 Gates Hall
Speaker: Sumit Gulwani, Microsoft
Title: Neuro-Symbolic AI: Making work, knowledge, and judgment compound
Host: Kevin Ellis
Abstract: General-purpose neural models can solve a remarkably broad range of problems. Yet much consequential work repeats. Reasoning from scratch each time adds cost, latency, and inconsistency. Neuro-symbolic AI offers another approach: neural models interpret intent, handle ambiguity and novel cases, and help author symbolic artifacts; those artifacts make recurring work more efficient, repeatable, inspectable, and auditable.
I will trace this idea from earlier program-synthesis systems to today's coding agents. Flash Fill synthesizes string transformations from examples, combining learned guidance with symbolic search, and remains far more efficient than frontier models in its domain. Coding agents can now lower the cost of building and maintaining such domain-specific neuro-symbolic tools for recurring tasks such as context-aware formatting and data cleaning, which higher-level agents can then invoke.
Coding agents can also distill repeated workflows into reusable neuro-symbolic scripts, enabling end users to create custom applications. They can turn natural-language grading criteria currently evaluated by LLM judges into executable symbolic evaluators with neural fallback. The symbolic artifact need not be code: structured knowledge bases can distill conversations and documents into provenance-aware knowledge, while rubrics can make recurring committee judgment inspectable and revisable. Across these forms, neural reasoning is amortized and prior effort compounds: work through scripts, knowledge through knowledge bases, and judgment through rubrics.
Bio: Sumit Gulwani is a Distinguished Scientist at Microsoft, where he connects ideas, connects people, and connects research with real-world practice. He is the inventor of Flash Fill, a popular AI feature in Excel that has also made its way into middle-school computing textbooks. His work spans agentic AI, AI-assisted programming, neuro-symbolic AI, and human-AI collaboration, with particular interests in spreadsheet intelligence and software engineering. He built PROSE, a blended research and engineering organization at Microsoft whose technologies have shipped across multiple Microsoft products. A strong believer in the power of storytelling, he sponsors internal programs at Microsoft to help technical teams communicate with greater clarity and impact. A passionate mentor, he also runs a thriving remote research fellowship program in India, nurturing young talent. His work has earned 15 paper awards, the Max Planck-Humboldt Medal, ACM SIGPLAN Robin Milner Young Researcher Award, and Fellowships from ACM and AAIA. He received his Ph.D. from UC Berkeley, where his dissertation was honored with the ACM SIGPLAN Outstanding Doctoral Dissertation Award. He earned his BTech from IIT Kanpur, where he was awarded the President’s Gold Medal and later the Distinguished Alumnus Award.
Website: https://www.linkedin.com/in/sumit-gulwani/
