New Paper: LM Harnesses Act as Compositional Generalizers, Not Crutches

RLanceMartin · x · 2026-11-21

a1zhang introduces the new paper "LM Harnesses are Compositional Generalizers": harnesses used to be a crutch for weak LLMs, but with stronger models from 2026 onward they become a mechanism for composing an LLM's existing and new capabilities to generalize well beyond training environments.

The paper explores how various harness abstractions enable LM systems to generalize over task length and task strategies.

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

Original post →

More from coding & agent

coding & agent channel →