Recursive Language Models: An All-in-One Deep Dive

Recursive Language Models (RLMs) are introduced by contrasting them with other agentic architectures like ReAct, CodeAct, and subagents. Using a simple task of generating word lists and counting letters, the limitations of each preceding method are demonstrated. For instance, ReAct is useful for narrow tasks with predefined tools, while CodeAct allows an LLM to generate its own tools, but both are prone to errors when reproducing results from memory. The subagent approach helps by dividing tasks but still requires loading entire outputs into the main agent's context, setting the stage for the problem that RLMs are designed to solve.