arXiv Computation and Language

Bridging Latent Reasoning and Target-Language Generation via Retrieval-Transition Heads

The paper investigates attention heads in multilingual Transformer models, distinguishing between retrieval heads that pull information from context and a newly identified class called Retrieval‑Transition heads (RTH) that direct the model toward a specific target language. Experiments across four multilingual benchmarks and two model families show that masking RTHs causes a larger performance drop than masking retrieval heads, indicating RTHs are crucial for chain‑of‑thought reasoning in multilingual LLMs. The study thus clarifies which attention heads are responsible for mapping to target languages, advancing our understanding of multilingual language models.

arXiv Computation and Language
Oct 5

Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation

The study explores how the language used for reasoning affects retrieval‑augmented generation (RAG) in a monolingual German setting. Using a German RAG question‑answering testbed based on the tabletop game The Dark Eye, the authors show that aligning the reasoning language with the query and retrieved documents improves performance, with German reasoning outperforming French reasoning. However, German reasoning still does not surpass the model’s native English reasoning, indicating that native multilingual reasoning is necessary for optimal results.

By Oliver Hauck, Mario Sanz-Guerrero, Katharina von der Wense
arXiv AI
Oct 26

Head Dimension Matters in Language Models for Long-Context Retrieval

The paper investigates how the dimension of attention heads influences long‑context retrieval in language models. Experiments show that while pretraining loss and short‑context tasks are largely unaffected, wider heads significantly improve retrieval performance at fixed model sizes. Ablation studies indicate that this benefit persists even when attention weights are frozen, suggesting that wider heads enhance adaptation of other model components such as embeddings and MLPs.

By Seiya Tokui, Shoichiro Yamaguchi, Sosuke Kobayashi
arXiv Computation and Language
Aug 27

Rethinking the Multilingual Reasoning Gap with Layer Swap

The study investigates the performance gap between native-language reasoning and English-pivoted reasoning in large language models. By creating extensive multilingual reasoning datasets and fine‑tuning specialists on Qwen/Qwen3-8B-Base, the authors find that the native reasoning gap is much smaller (1.9–3.5%) than previously reported. They analyze weight‑space changes, discover a language‑agnostic reasoning core in the middle layers, and propose a Layer Swap technique that transfers these mid‑layer updates from an English specialist to native specialists, effectively closing most of the gap while maintaining native chain‑of‑thought output.

By Maxence Lasbordes, Am\'elie Chatelain, Djam\'e Seddah
arXiv AI
Aug 12

From Reasoning Depth to Reasoning Breadth: Evaluating Multi-Point Associative Reasoning in Large Language Models

arXiv:2608. 10444v1 Announce Type: cross Abstract: Large language models (LLMs) have made substantial progress on reasoning tasks that require increasingly long and complex inferential chains.

By Si'an Xie (Beijing University of Posts and Telecommunications), Jiaxun Liu (Peking University), Biao Yang (Kuaishou Technology), Wei Yuan (Kuaishou Technology), Fan Yang (Kuaishou Technology), Tingting Gao (Kuaishou Technology), Ming Wu (Beijing University of Posts and Telecommunications)
arXiv Computation and Language
Oct 22

Task-Preserving Multilingual Instruction Tuning for Open-Ended QA in WMT 2026 MIST

The paper reports a submission to the WMT 2026 Multilingual Instruction Shared Task, where the authors fine‑tune Qwen3‑8B with LoRA on organiser‑provided data, EuroAlpaca, and CrossEuroAlpaca. Their system focuses on open‑ended question answering in ten target languages and shows a 7.8 % relative gain in Gemma‑4 judge score for target‑language QA‑OEG, as well as 41.5 % macro‑averaged ROUGE‑L and 2.9 % BERTScore‑F1 improvements on the Aya Evaluation Suite. However, the gains are task‑ and language‑specific, with some declines in context‑grounded QA and non‑target languages.

By Aleix Sant, Jordi Luque, Carlos Escolano