RLM: Recursive Language Models for Large Codebases
summary
Claims that a repo loads into a programmable REPL where the model writes inspection code and recursively delegates subquestions via llm_query; an independent reference implementation called RLM Code runs end to end on local and cloud models (demo uses Docker and Gemini) and produces an inspectable trajectory. The run makes two tool calls, emits JSONL traces and returns a final synthesized answer, showing the pattern works for large monorepos rather than just promising examples.
AI-written summary · about
chapters (12)
- 0:00RLM introduction
- 2:05Core thesis
- 3:07Lead engineer analogy
- 5:58Codebases vs plain text
- 6:19RLM Code purpose
- 7:51RLM implementation
- 9:44Live demo setup
- 12:11Coding agent demo
- 14:20Demo results summary
- 15:01Real-world use cases
- 16:05Where RLM concepts are already used
- 17:09Closing & Q&A
related talks
18mReductive Analysis with Compiler-Guided Large Language Models for Input-Centric Code(…)
19mLLM codegen fails and how to stop 'em
19mRL for Autonomous Coding
10mModel innovation & evaluation with model partners Q&A
47mHarnessing LLMs for Next-Level Security Analysis
16mThe Modular Imperative: Rethinking LLMs for Maintainable Software
25mEvaluating An LLM's Ability To Code
29mHow to roll your own LLM