ilo as the sandbox language inside Recursive Language Models (RLM).
Early. Phase 0 (using DSPy's RLM module with Python) is being explored privately to capture a token-cost baseline. Phase 1 (this repo) will implement an IloSandbox that lets a Recursive Language Model run ilo code instead of Python in its inner loop, with sub_llm() exposed as an ilo builtin.
Recursive Language Models give an LLM a sandbox where it can call sub_llm() recursively to reason over very large contexts. The original work uses Python as the sandbox language. Every recursive sub-call carries the sandbox source through the model's context, so the source language matters: fewer tokens per sub-call means cheaper recursion.
ilo is ~3× denser than Python on source tokens (measured across a 5-task suite). Swapping the sandbox from Python to ilo should reduce total RLM token cost on real tasks. This repo will test that hypothesis with measurements.
IloSandbox: anRLM-compatible sandbox that executes ilo codesub_llm(...): ilo builtin that calls back into the parent LLM- A small benchmark harness that runs the same task through Python-RLM and ilo-RLM and compares total tokens
- A writeup of the result, whichever way it goes
Repo created with stub README and LICENSE. Implementation has not started yet. Phase 0 (Python baseline) runs first to confirm the pattern before Phase 1 builds the ilo sandbox.
- Recursive Language Models paper: arxiv.org/abs/2512.24601
- Official RLM library: github.com/alexzhang13/rlm
- DSPy RLM module: dspy.ai/api/modules/RLM/
- The ilo language: github.com/ilo-lang/ilo
MIT. See LICENSE.