A model trained only on short tasks solves tasks 32 times longer

Claim

New research argues the harness that loops a model through subtasks is itself a generalization engine: train a Recursive Language Model only on short tasks and it still solves held-out tasks 8 to 32 times longer, transferring across domains far better than fine-tuning the transformer directly.

Reported metric: 8-32x longer tasks

Reporting date: 2026-07-21 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

• Welcome to July 21, 2026 (theinnermostloop.substack.com)