It’s Tokens All The Way Down: How RLMs are Different — Kevin Madura, AlixPartners
RLMs keep large inputs in a programmable environment, letting a model inspect them with code and delegate subsets to submodels instead of attending to every token at once.
A recursive language model treats context as a variable inside a Python-style REPL, writes code against it, and can delegate selected work to another model. The talk contrasts this with asking a base model to find **12 numbers across 30,000 tokens** directly.
Consider the pattern for large logs, data frames, repositories, or other inputs that benefit from deterministic slicing and computation. Define typed inputs and outputs, cap iterations, and return only relevant intermediate results to the main model.
A recursive language model treats context as a variable inside a Python-style REPL, writes code against it, and can delegate selected work to another model. The talk contrasts this with asking a base model to find **12 numbers across 30,000 tokens** directly. Consider the pattern for large logs, data frames, repositories, or other inputs that benefit from deterministic slicing and computation. Define typed inputs and outputs, cap iterations, and return only relevant intermediate results to the main model. The cited benchmark rose from **2.6% to 45.4% accuracy**, but the coding-agent comparisons were described as preliminary and potentially unfair. Cost, stopping behavior, sandboxing, and benefits over a carefully built conventional agent still require task-specific tests.
This introduces programmatic context manipulation as an alternative to placing a large input directly in a model or distributing it across a fixed agent chain. Deterministic slicing and computation can reserve model attention for selected evidence, with delegation kept optional and bounded. The benchmark supports testing the pattern, but preliminary comparisons leave its advantage over a carefully engineered conventional agent unresolved.