MIT CSAIL Explains: Recursive Language Models

Alex Zhang, Graduate Student | MIT EECS CSAIL https://alexzhang13.github.io/blog/2025/rlm/ DIrector: Rachel Gordon Preditor: Tim Malieckal Script Supervisor: Alex Shipps Chapters: 00:00 - Introduction 00:27 - LMs are more capable than we’ve realized 01:27 - What is a Recursive Language Model? 02:41 - What does your approach do that people are missing? 04:00 - What does treating a prompt as an environment enable? 05:36 - How does a RLM know what is worth looking at, and what isn’t? 07:30 - What does it take to build a RLM? 08:29 - An example of what a RLM can achieve that was out of reach before? 10:09 - Explain the ‘mismanaged genius’ hypothesis 11:33 - How does a RLM keep errors from compounding over many steps? 12:28 - How do RLMs fit into a multi-agent landscape?