New RL harness work says math training can improve essay writing too

iamrobotbear · x · 2026-07-21

The post argues that RLMs became exciting because they turned a hard non-coding problem—long-context reasoning—into a coding-like problem by using a harness to explore context.
It then highlights new work by Alex Zhang and collaborators: training RL on verifiable tasks such as math can improve both the verifiable task and a structurally similar non-verifiable task like essay writing.
The key claim is that a well-designed harness can induce generalization by composing task trajectories into shared structures, so the model does not need extra built-in generalization to transfer abilities.
The quoted paper language suggests a theory where harnesses create equivalence classes over trajectories, letting the same underlying trajectory solve tasks that look different on the surface.

More from AGI Musings