RLMs recurse into themselves: tail-recursive gated chains turn harnesses into models

a1zhang · x · 2026-11-20

a1zhang describes another experiment: harnesses that recursively invoke harnesses preserve the locally in-distribution property, with the harness abstracted as a sub-LLM — the origin of the Recursive Language Models (RLMs) name. They form meta-tasks where the RLM must solve each task in sequence to unlock the next, trainable via tail recursion over the harness itself.

Related event: MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches(10 posts)→

Original post →

More from coding & agent

coding & agent channel →