Recursive Language Model

Manages its own context through a persistent Python REPL, inspecting and transforming data without human oversight. The RLM was introduced by Alex L. Zhang of MIT CSAIL as a blog post in October 2025 and formalized with Tim Kraska and Omar Khattab in the paper Recursive Language Models (arXiv 2512.24601, 31 December 2025), with code released at github.com/alexzhang13/rlm; Prime Intellect made it a central research focus in its 1 January 2026 post. In the corpus it is the system behind a model taking over its own context and handling contexts two orders of magnitude beyond its window, the scaffold-level counterpart to the weight-level end-to-end test-time training of the same week.