← All works
Full-stack recurrence

Cross-Loop Query Handoff

Cross-Loop Query Handoff: How Later Loops Redirect Two-Hop Retrieval in the Ouro Looped Language Models

Ziyu Mu, Zehua Wang

A mechanistic study of how later loops in Ouro repair two-hop retrieval. Transplanting loop-3 attention queries into loop 1 rescues most questions that loop 1 got wrong, while late keys and values help far less, pointing to queries as the carrier of the later-loop correction.

Inside the method

Loop-3 queries → transplanted into Loop 1 (same layer, head, position)

Simplified conceptual schematic. Consult the paper for the complete architecture.

Recurrence family
Full-stack recurrence
Depth control
Analysis of released Ouro-1.4B and Ouro-2.6B
KV / state strategy
Late keys and values rescue far fewer cases than late queries

Reading note

On synthetic two-hop scenes, late queries rescue 90% and 97% of loop-1 failures. The effect is smaller on natural questions and replicates only in Ouro-2.6B; the paper does not locate where the bridge entity is computed.

Sources checked 2026-10-28. This catalog does not imply independent reproduction.

Related works

Cite this work

@misc{crossloopqueryhandoff2026,
  title = {Cross-Loop Query Handoff: How Later Loops Redirect Two-Hop Retrieval in the Ouro Looped Language Models},
  author = {Ziyu Mu and Zehua Wang},
  year = {2026},
  eprint = {2610.34639},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34639}
}

Suggest a correction on GitHub ↗