Skip to content

Recursive Language Models Explained: Moving Long Context into an External Environment

For / Key Points

This is for readers who want a view of long-context handling beyond RAG or simply widening the context window. It visualizes how RLMs keep long text in an external environment and inspect, split, and recursively process it through code.

Text Summary

The paper Recursive Language Models was submitted to arXiv on December 31, 2025 and revised as v2 on January 28, 2026.1 Its core move is not simply to extend the context window. It changes where the long context lives: outside the model, in an environment the model can programmatically inspect, split, and recursively process.

  • RLMs do not put the whole long prompt directly inside the model context window.
  • The long prompt is treated as a variable in an external environment, which the model can inspect through code.
  • The paper reports that RLMs outperform vanilla frontier LLMs and common long-context scaffolds on several long-context tasks.

Open the full-screen RLM interactive explainer if you want the JSX article without the blog wrapper.

Summary

RLMs are best read as a shift in where an LLM's working memory lives. Instead of pushing long text into the context window, the system keeps it in an external environment and lets the model inspect, split, recursively process, and aggregate it through code.

That makes the paper relevant not only to long-context benchmarks, but also to future agent architectures and document-processing systems.

  1. Alex L. Zhang, Tim Kraska, Omar Khattab, Recursive Language Models, arXiv:2512.24601. ↩