Recursive Language Models

What we'll be reading
This session, we'll dive into Recursive Language Models (Zhang, Kraska & Khattab — MIT CSAIL), a compelling new inference paradigm that lets LLMs process inputs far beyond their native context windows.
Paper link: https://arxiv.org/pdf/2512.24601
Why this paper?
Even frontier models like GPT-5 suffer from context rot — performance degrades sharply as inputs get longer and tasks get more complex. RLMs propose an elegant solution inspired by out-of-core algorithms: instead of feeding the entire prompt into the transformer, the model treats the prompt as an object in an external environment (a Python REPL) and writes code to peek into, decompose, and recursively call itself over snippets of the input.
The results are striking: RLMs handle inputs up to two orders of magnitude beyond the base model's context window and consistently outperform direct calls, context compaction, retrieval agents, and code-generation agents — often by double-digit margins — across tasks ranging from needle-in-a-haystack to quadratic-complexity pairwise reasoning.
Outline

• Introduction
• Recursive Language Models
• Scaling Long Context Tasks
• Results and Discussion
• Limitations, Future Work and Conclusion

How to prepare
Give the paper a read beforehand. Come with questions, hot takes, or related work to share!
Who is this for?
Anyone curious about LLMs, inference-time scaling, or long-context processing. All levels welcome — we'll make sure to cover the key concepts together.
See you there!