Reproduce “Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seunghan Kim, Minyeong Choe, Hyunil Kim, Haehyun Cho
The authors identify attention-head pools involved in multilingual two-hop reasoning and test whether amplifying them can recover answers. A general pool routes bridge entities across languages, alongside language-specific pools with little overlap. In two models, ablating the general pool increased gold-answer NLL 39–89 times more than random-head ablation. Amplification rescued 51.7% of selected Japanese failures, not overall accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.