Recursive Language Models

the repro published: https://github.com/alexzhang13/rlm

We’ve all been sold the dream of the million-token context window—the idea that you can dump entire libraries into an AI and it will perfectly understand everything. But here is the dirty secret researchers are finally admitting: it does not really work. As you fill up a model’s memory, it actually gets dumber and starts forgetting the middle parts, a phenomenon known as context rot. Enter Recursive Language Models, a new paper from MIT researchers that claims to solve this not by making the model's brain bigger, but by giving it a better strategy. Instead of trying to memorize a 1,000-page book in one glance, these models treat the text like a database that they can query intelligently.

Think of a standard AI as a student trying to cram for an exam by reading the textbook cover-to-cover in one sitting; eventually, their brain turns to mush. A recursive model acts more like a project manager with a laptop. The massive text isn't shoved into the model's immediate focus. Instead, it sits in an external environment. The model then writes code to peek at specific sections, search for keywords, or even spawn sub-agents to go read specific chapters and report back with summaries. It breaks one giant, impossible task into tiny, manageable chunks of code and reading.

This is a massive vibe shift for 2026. For years, the AI race has been about brute force—building bigger GPUs to handle bigger inputs. This paper suggests we’ve been doing it wrong. By using this recursive method, a much smaller and cheaper model was able to outperform the massive GPT-5 on heavy reasoning tasks. It implies that the future of AI isn't about building a god-like super-brain that knows everything at once, but rather building a smart agent that knows how to look things up. Expect a lot of debate on whether context windows are dead and if recursion is the new kingmaker.

paper: https://arxiv.org/pdf/2512.24601