Yu Su is co-founder and chief executive of NeoCognition, which develops AI agents that acquire specialized expertise through experience. He is also an associate professor and Innovation Scholar at The Ohio State University, where his research addresses a central limitation of contemporary AI: powerful models often lack the accumulated judgment needed to operate reliably in unfamiliar organizations and digital environments.
Su studied computer science at Tsinghua University and earned his doctorate in the same field at the University of California, Santa Barbara. He subsequently worked as a senior researcher at Microsoft Semantic Machines before joining Ohio State, where he co-directs the natural language processing group and leads research across the ICICLE AI Institute and Imageomics Institute. He became an associate professor in 2025 and received a Sloan Research Fellowship that year.
His research has helped establish several building blocks for capable agents. Mind2Web, introduced in 2023, tests agents against more than 2,000 tasks across 137 real websites and 31 domains. SeeAct investigates how multimodal models translate website understanding into executable actions. His other collaborations include MMMU, a benchmark for knowledge-intensive multimodal reasoning, and HippoRAG, a hippocampus-inspired memory architecture integrating language models, knowledge graphs, and graph-based retrieval.
Su founded NeoCognition with fellow Mind2Web researchers Xiang Deng and Yu Gu. The company emerged from stealth in 2026 with $40 million in seed financing, applying research on agents and continual learning to organization-specific expertise.
- Intelligence is not expertise. General intelligence helps an agent reason through unfamiliar problems; expertise supplies the right context, recognizes exceptions, narrows the search for solutions, and develops practical judgment through accumulated experience.
- Digital work consists of specialized micro-worlds. Coding offers structured languages, explicit tests, and clear feedback. Ordinary organizations present idiosyncratic software configurations, local rules, and institutional knowledge, creating a modern version of Moravec’s paradox: sophisticated symbolic reasoning can coexist with unreliable everyday computer use.
- Continual learning creates reusable expertise. Su treats experience as material that agents can compress into model parameters, external memory, vector representations, graphs, and reusable skills. His research questions include combining parametric and nonparametric learning while balancing adaptability against reliability.
- Better learning can outperform bigger models. Once general intelligence reaches a sufficient threshold, Su argues, improved learning mechanisms could produce increasingly deep expertise without indefinitely scaling the underlying model. Organization-specific human–AI learning loops could accumulate institutional memory and eventually expand access to specialized support in healthcare, finance, and education.
