Metraipolis › archive › dc_f73f2b600e41

Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs — arXiv:2610.09733

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.09733
License (as archived)
CC0-1.0
Archived
2026-10-08T11:26:14.087Z
SHA-256 of the archived text
f06f0d25cafd6bc0fec79bdd15f886d8e8128adf65d9455a69b6e87ede9ff947
Stable URL of this page
https://metraipolis.com/documento/dc_f73f2b600e41

How to cite

Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs — arXiv:2610.09733 — arXiv (en), https://arxiv.org/abs/2610.09733 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_f73f2b600e41 (Metraipolis archive dc_f73f2b600e41, sha256 f06f0d25cafd6bc0fec79bdd15f886d8e8128adf65d9455a69b6e87ede9ff947)
BibTeX
@misc{dcf73f2b600e41,
  title = {Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs — arXiv:2610.09733},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.09733},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_f73f2b600e41; sha256 f06f0d25cafd6bc0fec79bdd15f886d8e8128adf65d9455a69b6e87ede9ff947},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.09733. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Multilingual LLMs answer the same multi-hop reasoning question across languages, but we lack a mechanistic account of whether they share an internal circuit. We identify Bridge Routing Heads (BRH) in two large multilingual LLMs through a three-stage pipeline. The resulting language-specific head sets exhibit near-complete mutual exclusivity across the five languages, with a mean Jaccard similarity of only 0.017 for Llama 3.1 70B and 0.057 for Qwen 2.5 72B, revealing language-idiosyncratic circuits. Ablating general BRH increases two-hop Negative Log-Likelihood (NLL) by 39-89x the random-head baseline, providing direct causal evidence of their role. Amplifying these heads in a failing target-language pass rescues up to 51.7% of cross-lingual failures, with no training. The two models share this dual-circuit pattern but allocate heads differently: Llama concentrates chaining in a large general pool, while Qwen leans on larger language-specific pools. Together these results show that activation-level intervention alone can recover correct answers from cross-lingual reasoning failures.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2610.09733 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-08T10:24:06.334Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).