Scaling LLM context windows to millions of tokens usually leads to reasoning degradation. This paper introduces Recursive Language Models to process prompts up to 10M tokens by treating them as an external symbolic environment. Modern frontier models suffer from context rot. Even when a model possesses a large physical context window, its ability to reason over that information drops significantly as the prompt grows. Existing solutions like context compaction or summarization attempt to compress the input. These methods often fail on complex tasks because they discard fine-grained details that might be necessary for the final answer. Standard LLMs attempt to ingest the entire prompt into their neural attention mechanism. This forces the model to process every token simultaneously, which leads to noise and memory limitations. Recursive Language Models (RLMs) move the prompt out of the model and into a symbolic workspace. The model interacts with the prompt through a programming environment rather than direct token ingestion. The RLM process follows a specific execution loop: → The input prompt is loaded as a variable into a Python environment. → The model writes code to search, slice, or filter the prompt. → The model programmatically creates sub-tasks based on what it finds. → It recursively calls versions of itself to solve these smaller snippets. → The results are stored as variables and stitched together to form a final response. This method is distinct from prior agentic scaffolds. While previous work focused on breaking down the steps of a task, RLMs allow the input data itself to scale by offloading it to a variable in memory. By treating the prompt as an object to be manipulated, the model can navigate 10M+ tokens without ever having to "read" the entire document into its own context window at once. The primary risk is trajectory variance. Because the model chooses its own path through the data, it can occasionally perform redundant sub-calls or get stuck in verification loops, which increases runtime. The authors address this by using a system prompt that encourages efficient chunking. They also find that using a smaller, cheaper model for recursive sub-calls maintains a high performance-to-cost ratio. The researchers tested RLMs using GPT-5 and Qwen3-Coder across four difficult long-context benchmarks. The results show significant improvements over base models and standard retrieval methods: → RLMs handled inputs two orders of magnitude larger than model context limits. → 91.3% accuracy on BrowseComp-Plus compared to 0% for the base model. → 58% F1 score on complex pairwise reasoning tasks where base models failed. ✓ Effective scaling to 10M+ tokens ✓ Reduced context rot ✓ Comparable or lower inference costs This research suggests that the path to infinite context windows may not be through architectural changes alone. Instead, it demonstrates that context management can be treated as an inference-time reasoning task. RLMs prove that models can handle professional-scale datasets by using symbolic tools to manage their own attention. This shifts the focus from building larger neural memories to developing models that can programmatically navigate external information.