Recursive Language Model

MIT just demonstrated that Recursive Language Models (RLMs) can process inputs 100x longer than GPT-5 — without degrading performance. Instead of squeezing everything into a fixed context window, RLMs treat text like a searchable external system, recursively querying and refining what matters. The outcome: → GPT-5 accuracy collapses as input scales → RLM(GPT-5) sustains 60–100% accuracy beyond 1M+ tokens No context loss. No forced compression. Entire books. Massive repositories. Multi-hour transcripts. Same underlying model. Comparable cost. Exponentially stronger long-context reasoning. The context-window era is ending. The recursive era begins. 🔁