Language-model harnesses can drive compositional generalization

a1zhang · x · 2026-07-20

A new paper argues that well-designed language-model harnesses can do the generalization work that transformers themselves struggle with.

The authors observe that in training RLMs, structurally similar tasks can collapse to nearly identical trajectories.
They describe harnesses as inducing a quotient over task trajectories, letting different-looking tasks become token-for-token similar.
In experiments, models trained only on short tasks generalized to unseen tasks that were 8–32× longer.
They also show cross-domain transfer when tasks share a decomposition strategy, such as transferring from author-matching in essays to math-problem grouping.
The takeaway: harness design may reduce data/rollout cost while expanding the set of tasks solvable in post-training.

More from coding & agent