Recursive Language Models

Recursive Language Models (RLMs) introduce a general inference strategy that allows Large Language Models (LLMs) to process arbitrarily long prompts (exceeding 10 million tokens) by treating them as external environment variables within a Python REPL. The model programmatically examines, decomposes, and recursively calls itself over prompt snippets. This approach achieves superior performance and robustness on diverse long-context tasks compared to direct LLM calls and other scaling methods, mitigating "context rot" and making LLMs effective for applications requiring extensive context understanding.