Recursive Language Models: Scaling infinitely beyond the Context Window

Recursive Language Models (RLMs) are a novel inference framework developed by MIT CSAIL. Unlike traditional methods that feed entire prompts into a model, RLMs treat the context as an external variable within a Python REPL environment, allowing the AI to explore, partition, and summarise data programmatically. This strategy effectively scales to near-infinite context lengths by utilising recursive sub-calls to process manageable chunks, thereby preventing the “context rot” typically seen in long-form tasks. Empirical results demonstrate that RLMs significantly outperform frontier models, such as GPT-5, on complex benchmarks while maintaining comparable or even lower API costs. Beyond benchmarking, the sources highlight broader research background in GPU programming, software engineering evaluation, and the potential for RLMs to revolutionise agentic workflows in production environments. Overall, RLMs are a scalable, model-agnostic alternative to human-engineered agent scaffolds for high-precision, long-context reasoning.