MIT PhD claims harness design can make RLMs generalize from short tasks to 100× longer ones

a1zhang · x · 2026-07-21

MIT PhD student Alex Zhang argues that the harness can do much of the generalization work in recursive language models (RLMs): if the surrounding design is clever, some scaling gains can be had “for free.”
The key observation is that models trained only on short tasks can transfer surprisingly well to much longer problems in a similar domain, sometimes at 100× longer context lengths.
In the quoted explanation, the claim is that a well-designed harness effectively induces structural equivalence between different trajectories, so the transformer does not need to learn all the generalization itself.
The post frames this as a shift in where generalization lives: less in the base model, more in the orchestration/composition layer.

More from Research