In this presentation we present Recursive Language Models (RLMs), developed by MIT CSAIL. RLMs solve the problem of context rot in long inputs by moving the full context outside the model into a Python REPL environment. The main LLM writes code to read, split, search, summarize, and recursively call smaller or cheaper instances of itself on smaller chunks. This keeps the root model’s context small while handling very large inputs.

RLMs work with any existing LLM and scale effectively to 10 million+ tokens without retraining or losing performance. On benchmarks like OOLONG and BrowseComp-Plus, RLMs clearly outperform standard frontier models and common long-context methods, often at similar or lower cost.

We show a simple PyTorch implementation and introduce RLM-Qwen3-8B, a post-trained model that learns native recursion and performs much better than its base on long-context tasks. RLMs offer a practical way to build agentic and deep-research systems today.