MIT researchers unveiled a paper that reimagines how large language models handle massive text. Instead of expanding token windows—an arms race where Gemini, GPT series and Llama chase millions of tokens—MIT proposes Recursive Language Models (RLM). The new architecture treats the entire input as a mutable variable inside a Python REPL, letting the model slice, search and call itself recursively.

By converting context from a static string to an interactive environment, RLM sidesteps the memory limits that plague traditional transformers and the retrieval bottlenecks of RAG pipelines. In a demo, the system answered a request to list the first 100 powers of two by loading the full dataset into a variable, inspecting its structure with regex, breaking the task into sub‑chunks, and invoking smaller model calls for each piece. Results were assembled and verified through multiple passes, producing a flawless, long‑form output.

The approach promises stronger reasoning on huge documents but adds latency and engineering complexity, especially around code execution safety. Early adopters may reserve RLM for tasks where accuracy outweighs speed.