Recursive LM harnesses generalize from short tasks to jobs 8–32× longer

heghbalz · x · 2026-07-21

A paper argues that language model harnesses can act as compositional generalizers.
The core idea is to use recursive LMs that offload context and call sub-agents, so the main model repeatedly sees a simple, reusable structure rather than one giant task.
The authors train on short tasks but report generalization to tasks 8–32× longer.
They say each step stays locally in-distribution, and the model learns the decomposition pattern behind the task instead of memorizing the task itself.
The framing shifts generalization from “just add more data and longer context” to a harness-design problem.

More from Research