Researchers argue the harness, not the transformer, should generalize through composition

eliebakouch · x · 2026-07-22

A short reply celebrating a research idea: transformers struggle to generalize to unseen tasks, so the proposed direction is for the harness itself to generalize through composition.

The post also points to an observed property in RLM training: for tasks with shared structure that look different on the surface, the root model can learn the same trajectory, effectively treating the two task traces as equivalent.

Related event: Research Suggests RLM Generalization is Driven by External Harness(10 posts)→

Original post →

More from coding & agent

coding & agent channel →