Recursive Language Models (RLMs)

Researchers have developed Recursive Language Models (RLMs), allowing AI models to break down extremely long text inputs by making recursive calls to themselves within a Python environment. RLMs enable a language model to treat input context as a variable it can examine, chunk, and query through sub-calls rather than processing everything at once. Testing showed RLMs using GPT-4o-mini outperformed regular GPT-4o on difficult long-context tasks while being cheaper per query, and successfully handled contexts exceeding 10 million tokens without performance degradation.