RLMs, or Recursive Language Models, address a significant issue in AI called context rot, where the quality of output deteriorates as the amount of context increases. The Gemini 2.5 paper highlighted that their performance declined beyond 100,000 tokens, despite a 1 million token limit. Context rot is not about capacity but quality; models continue to generate outputs, but their accuracy drops. RLMs tackle this by managing two types of context: tokenized, which fills the LLM’s window, and programmatic, which exists in a coding environment. This setup allows the model to filter and analyze extensive inputs efficiently without sacrificing performance. An example of RLMs in action involved processing over 400 megabytes of Stable Diffusion prompts. The model successfully identified common celebrities used in the prompts without crashing, unlike traditional LLMs that would fail with such large inputs. However, RLMs have limitations. They can be slow, as the task took several minutes with multiple calls to the model. Not all models perform well with RLMs; for instance, Qwen3-30B-A3B struggled with the task and ran out of time. RLMs also benefit from recent advancements in coding capabilities, making them effective at turning long context problems into coding challenges. Beyond mitigating context rot, RLMs can help discover and optimize problem-solving strategies. By analyzing traces of their operations, users can identify patterns and refine approaches for future tasks. Teams have found success using RLMs in large context scenarios, like coding across vast codebases or conducting research in extensive datasets. However, they are not suitable for smaller context problems, as the exploration may take longer than simply including the relevant context in the prompt. RLMs do not solve other issues like context poisoning, where inaccurate information could negatively impact results.