Recursive Language Model implemented, evaluated, explained
Recursive Language Model implemented, evaluated, explained
How do you process data larger than an LLM's context window? Instead of expanding context, expand the workspace. This video explains the Recursive Language Model (RLM) technique using a "cookie jar" analogy and shows benchmark results: 87-91% accuracy on 9 SCROLLS tests using only ~3000 tokens per iteration.
⚠️ AUDIO NOTE: I am aware of the audio problems; please turn on captions to help clarify what is being said.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
MORE IN THIS SERIES:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
▶ Can AI Find a Secret Hidden in War and Peace?
• Can AI Find a Secret Hidden in War and Peace?
▶ Custom Code in a Sandbox? RLM and WASM
• Custom Sandbox Code ? RLM and WASM
▶ Why I Let an LLM Compile Native Binaries
• Why I Let an LLM Compile Native Binaries
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
BACKGROUND
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
This is my Rust implementation of RLM, not the Python implementation by the paper's authors. I built this from scratch using Rust, DeepSeek via LiteLLM, and vibe coding with Claude.
CAPABILITY LEVELS (Roadmap):
• L1 (DSL): Built-in commands for text operations (find, regex, count, filter, extract)
• L2 (WASM): LLM generates Rust code → compiled to WebAssembly sandbox
• L3 (CLI): LLM generates Rust code → compiled to native binary for large datasets
• L4 (LLM): Recursive delegation - LLM delegates chunks to sub-LLMs for semantic analysis
LINKS:
📄 Paper: https://arxiv.org/abs/2512.24601
💻 Code: https://github.com/softwarewrighter/r...
More from Software Wrighter Lab:
Blog: https://software-wrighter-lab.github.io/
Discord: / discord
GitHub: https://github.com/softwarewrighter
#RLM #RecursiveLanguageModel #Rust #VibeCoding #Deepseek #LiteLLM #LLM #AI #MachineLearning #SCROLLS #Benchmark #ContextWindow #NLP #Programming #SoftwareEngineering #OpenSource #AITools #TechDemo