The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The paper investigates multilingual confidence calibration in large language models, revealing that non‑English languages are systematically less well calibrated than English. By analyzing internal representations, the authors find that late‑intermediate layers provide a more reliable confidence signal than the final layer, which is biased by English‑centric training. They propose training‑free methods such as Language‑Aware Confidence Ensemble (LACE) to adaptively select optimal layers per language, aiming to improve global equity and trustworthiness of LLMs.
The paper challenges the assumption that improving cross‑lingual alignment automatically enhances cross‑lingual transfer. Using XLM‑R models aligned on token, sentence, and masked‑language‑modeling objectives across four language pairs, the authors evaluate zero‑shot transfer on part‑of‑speech tagging and sentence classification. They find that embedding‑based alignment metrics poorly predict downstream performance and that alignment and task gradients are often nearly orthogonal, especially when operating at different representational levels.
The paper proposes the interlingua hypothesis, suggesting that large language models translate by encoding a source sentence into a latent, task‑agnostic feature space and then decoding a target sentence from that space. Three lines of evidence support this: (1) BLEU variance across language pairs is largely explained by language‑specific competences without pair‑specific interactions; (2) many model components influence both monolingual and translation tasks; and (3) fine‑tuning on monolingual data recovers most translation gains seen with aligned documents. These findings converge to support the hypothesis and point toward new ways to understand and improve LLM translation.
The paper proposes a multilingual adaptation of an English quality classifier to select high‑quality training data for large language model pretraining across more than 100 languages. The method trains a small multi‑layer perceptron on Transformer embeddings, using machine‑translated text scored by the English classifier as labels. Experiments on 1B, 3B, and 8B parameter models show that the adapted classifier preserves benchmark performance and regional knowledge, and it generalizes to unseen languages by learning the original scoring criteria.
The paper investigates how to fairly compare language models across languages, noting that current evaluation methods vary widely and lack empirical validation. By training controlled monolingual models on parallel data and testing multilingual LLMs, the authors find that many normalized metrics suffer from biases due to tokenization, encoding, and orthographic differences. Instead, they recommend using sentence‑level negative log‑likelihood over semantically equivalent sequences for more reliable cross‑lingual comparisons.
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models. "whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."