Yu Su

Su’s early work studied natural-language interfaces to data. He held a visiting appointment at Ohio State and worked at Microsoft Semantic Machines before starting his Ohio State assistant professorship in January 2020. His own news records promotion to associate professor in May 2025. He now also cofounded NeoCognition with Mind2Web collaborators Xiang Deng and Yu Gu.[1][2][3]

Retrieve connected knowledge without retraining the model

HippoRAG 2 combines language-model processing with knowledge structure and Personalized PageRank, a graph-ranking method seeded by the query. Deeper passage integration and more effective online model use aim to preserve ordinary factual retrieval while improving associative and multi-hop questions. The paper evaluates these memory tasks as a form of non-parametric continual learning: new knowledge enters an external retrieval system, without changing the language model’s weights.[8]

Evaluate web agents across unfamiliar sites

Mind2Web supplies natural-language tasks and recorded action sequences from real websites across many domains. Its baseline filters large HTML pages with a smaller language model before asking a larger model to choose an action. The dataset tests whether an agent can transfer to new tasks, websites and domains rather than succeeding only in a simplified environment it already knows.[10]

In his continual-learning talk, Su distinguishes reasoning through an unfamiliar problem from acting efficiently with accumulated domain knowledge. Organizations have local software, rules and exceptions that an agent must learn through experience. He proposes treating them as specialized micro-worlds, with feedback and memory helping the agent develop an increasingly useful model of its environment. This frames expertise as an ongoing learning process instead of a static prompt containing every possible situation.[7]