Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.09733

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

Seunghan Kim·Minyeong Choe·Hyunil Kim·Haehyun Cho
View original source ↗Read source document ↗Export metadata ↓

Abstract

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.

Explore related subjects

cs.CL

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Seunghan Kim, Minyeong Choe, Hyunil Kim, Haehyun Cho. 2026-10-07. Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs. https://arxiv.org/abs/2610.09733

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Enabling Quantum Natural Language Processing for Hindi Language

Quantum Natural Language Processing (QNLP) is taking huge leaps in solving the shortcomings of classical Natural Language Processing (NLP) techniques and moving towards a more "Explainable" NLP system. The current literature around QNLP focuses primarily on implementing QNLP techniques in sentences in the English language. In this paper, we propose to enable the QNLP approach to HINDI, which is the third most spoken language in South Asia. We present the process of building the parameterized quantum circuits required to undertake QNLP on Hindi sentences. We use the pregroup representation of Hindi and the DisCoCat framework to draw sentence diagrams. Later, we translate these diagrams to Parameterised Quantum Circuits based on Instantaneous Quantum Polynomial (IQP) style ansatz. Using these parameterized quantum circuits allows one to train grammar and topic-aware sentence classifiers for the Hindi Language.

cs.CL↗
arXiv2026

Foundations of Large Language Models

This is a book about large language models. As indicated by the title, it primarily focuses on foundational concepts rather than comprehensive coverage of all cutting-edge technologies. The book is structured into six main chapters, each exploring a key area: pre-training, generative models, prompting, alignment, inference, and reasoning. It is intended for college students, professionals, and practitioners in natural language processing and related fields, and can serve as a reference for anyone interested in large language models.

cs.CL↗
arXiv2026

Fair-GPTQ: Bias-Aware Quantization for Large Language Models

The high memory demands of generative language models have drawn attention to quantization, which reduces memory usage by mapping model weights to lower-precision integers. However, recent empirical studies show that, while efficient, quantization can increase the likelihood of generating biased outputs and degrade performance on fairness benchmarks. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization.

cs.CL↗