The paper investigates how multilingual large language models can unintentionally switch languages during generation. It compares three techniques—ValSel, FreqSel, and AnnSel—for pinpointing latent variables that control language choice in cross‑layer transcoders. Using new multilingual benchmarks and targeted interventions on Gemma‑2‑2B and Qwen3‑4B, the study finds all methods can steer output language, with FreqSel performing best and AnnSel providing interpretable selections via explicit annotations.
By Ryo Mitsuhashi, Sabri Boughorbel, Majd Hawasly
arXiv:2601.16390v2 Announce Type: replace-cross
Abstract: Large language models exhibit strong multilingual capabilities, yet significant performance gaps persist between dominant and non-dominant la...
By Rhitabrat Pokharel, Ameeta Agrawal, Tanay Nagar
arXiv:2604. 03532v2 Announce Type: replace-cross Abstract: Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult.
By Sing Hieng Wong, Hassan Sajjad, A. B. Siddique
The paper investigates whether training small decoder-only transformers on code‑switched text can induce cross‑lingual alignment. Using two 100‑million‑word multilingual corpora—one a mix of English, Dutch, and Chinese BabyBabelLM data, and another generated by inserting word‑ and sentence‑level code‑switching via an LLM—the authors find that code‑switched training aligns representations of parallel text, especially across different scripts, and that this alignment persists when later training on monolingual documents. A curriculum that progresses from word‑level code‑switching to sentence‑level code‑switching and finally to monolingual data yields models that outperform baselines on the BabyLM evaluation suite, demonstrating that code‑switching curriculum learning is an effective data augmentation strategy for multilingual pretraining.
By Dries Rooryck, Alex Cai, Yonatan Belinkov, David Alvarez-Melis, Kiant\'e Brantley
arXiv:2510. 05678v2 Announce Type: replace-cross Abstract: While large language models (LLMs) have achieved notable progress in multilingual settings, their performance remains uneven across languages as LLMs often rely on English-centric latent representations.
By Haneul Yoo, Jiho Jin, Kyunghyun Cho, Alice Oh
The paper addresses the challenge of identifying code‑switched utterances in language identification systems. It revisits the MaskLID approach, highlighting its overreliance on word‑level language association scores, and reformulates its optimization as an Integer Linear Program to incorporate clear, interpretable constraints. These enhancements significantly improve performance across ten diverse languages on code‑switching benchmarks, with the authors releasing code and data for reproducibility.
By Joanna Rado{\l}a, Josep Maria Crego, Fran\c{c}ois Yvon
The paper introduces a lightweight approach to detect code‑switching (CS) directly from speech by leveraging language‑selective hidden‑state dimensions in pretrained speech encoders. Using monolingual recordings to identify these dimensions, the method aggregates frame‑level activations to predict CS presence, achieving a mean F1 score of 0.798 on the CS‑FLEURS dataset with the Whisper‑small encoder. This demonstrates that pretrained speech models inherently encode CS information, offering a practical tool for expanding multilingual corpora.
By Jie Chi, Zakaria Aldeneh, Masha Fedzechkina, Maureen de Seyssel, Natalie Schluter
The paper investigates whether different latent language probes—GMM-based representation probes and decoding-based probes—measure the same phenomenon in multilingual language models. Across various model families, training regimes, domains, tasks, checkpoints, and up to 27 languages, the authors find systematic disagreement: representation probes indicate earlier cross‑lingual mixing, while decoding probes reveal sharper, English‑biased language signals. These differences correlate with model multilinguality and training progression but remain relatively stable across domains, suggesting that current probes capture distinct aspects of multilingual processing rather than a single internal lingua franca.
By Deniz Bayazit, Badr AlKhamissi, Antoine Bosselut
arXiv:2606.18389v2 Announce Type: replace
Abstract: Large language models (LLMs) have become an effective tool for synthetic data generation, including for low-resource languages, where generated dat...
By Jan Cegin, Daniil Gurgurov, Yusser Al Ghussin, Simon Ostermann
arXiv:2610.16601v1 Announce Type: cross
Abstract: Processing multilingual text requires both identifying a broad range of languages and distinguishing the language of each token in inputs containing...
By Seongtae Hong, Youngjoon Jang, Seungyoon Lee, Jungseob Lee, Heuiseok Lim
arXiv:2609.07037v1 Announce Type: new
Abstract: Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional...
By Takeru Hiramatsu, Kyohei Atarashi, Koh Takeuchi, Hisashi Kashima
The paper introduces a new method for aligning cross‑lingual representations in decoder‑only large language models by leveraging the outputs of mixture‑of‑experts (MoE) routers instead of traditional hidden‑state alignment losses. This approach pools router outputs over many tokens, providing more reliable sequence‑level comparisons across languages. Experiments on four open‑source MoEs show that the routing loss not only aligns hidden representations but also improves multilingual performance across a diverse evaluation suite.
By Lucas Bandarkar, Clark Peng, Ahmed Haj Ahmed, Aditi Khandelwal, Nanyun Peng