Summary
Kevin Madura of consulting firm AlixPartners introduces Recursive Language Models (RLMs) to an AI Engineer audience, explaining how this emerging paradigm differs fundamentally from both standard LLM tool calling and conventional agentic coding. Where a tool call passes a JSON string to an external program and returns a string, an RLM treats context as a symbolic object it can manipulate directly inside a Python REPL environment — and can recursively delegate sub-problems to other LLMs (including itself) running in that same environment.
Madura traces the concept to work by DSPy creator Omar Khattab and collaborator Alex, citing benchmarks where RLMs demonstrate measurable advantages: on the ULong benchmark for long-context question answering and BrowseComp, the RLM outperforms both standard models and a GPT-5 + BM25 tool-calling setup on the price-performance curve. A long-chain-of-thought benchmark shows accuracy jumping from 2.6% to 45.4% — a result driven by the RLM’s ability to write code, execute it in the REPL, and harvest computed results rather than attending over all tokens simultaneously.
The talk outlines where RLMs shine (documents exceeding context windows, logic and calculation tasks amenable to code, long-horizon exploration like tax-code analysis) and where they fall short (tasks that fit in context, latency-sensitive workloads, or situations where the base model is already a strong coder). Madura also contrasts RLMs against coding agents like Claude Code, arguing that for structured production workloads, RLMs offer more deterministic, composable behavior.
📺 Source: AI Engineer · Published September 09, 2026
🏷️ Format: Deep Dive
