Karthik Narasimhan
Princeton University
- Research field
- Language models that act in interactive environments, language agents
Karthik Narasimhan is a computer scientist at Princeton University whose group builds language models that act: answering questions by searching, carrying out tasks in text worlds and on websites, and fixing real software issues. In a 2022 paper with Shunyu Yao and colleagues at Google he showed that a model which alternates written reasoning with actions in an environment, a pattern the paper named ReAct, invents fewer facts than a model that only reasons and plans better than a model that only acts, and that combining the two gives the best results on question answering, fact checking, household tasks, and web shopping. His group also built the SWE-bench benchmark of real GitHub issues. The unit presents the ReAct loop as its key algorithm for agents and the paper's experiments in the research spotlight.
In the masterclass
Section 5, right after the ReAct loop and its drawing, presents the paper's comparison of reasoning only, acting only, and the two interleaved.
The research is presented in Unit 15: Reasoning, Tools, and Agents, in a research spotlight box that explains the result at the level of the unit and links it to what you have just learned.