RLM paper argues harnesses, not transformers, drive task generalization

a1zhang · x · 2026-07-21

The post highlights a section from an RLM paper arguing that a “harness” can generalize by composition, rather than relying on the transformer itself to generalize.
The core claim is that when training RLMs, trajectories for tasks with shared structure can become nearly identical in latent form, so the root model treats them as the same trajectory and transfers behavior across tasks.
The thread notes an “open secret” in frontier model training: even without direct test-set leakage, models may still be trained on lookalike tasks, which can make benchmark results look better than they really are.
The attached figures compare evaluation trajectories against nearest training trajectories using several distance metrics such as Levenshtein, n-gram containment, Jaccard, weighted Jaccard, and length ratio, showing that best RLM rollouts often stay much closer to training trajectories than a plain Transformer baseline.
The post’s broader point is that better harness design may be what induces transfer, and that current proxy metrics still fail to fully capture semantic similarity between trajectories.

More from Research