The Flow has not summarised this story yet — read it at arXiv Computation and Language.
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.
Manacá-1B is a 1.72‑billion‑parameter, open decoder‑only language model trained from scratch for Brazilian Portuguese, released with a fully containerized, reproducible training pipeline and complete logs. The authors evaluate it against nine open baselines on four Portuguese benchmarks, reporting standard errors and paired significance tests, and find that Manacá-1B outperforms smaller models on LAMBADA‑PT while remaining competitive on commonsense completion. They also uncover a tokenizer‑related evaluation pitfall that can drastically lower accuracy and provide a simple fix, releasing all code, logs, and corrected tokenizer for full reproducibility.
The paper investigates how many languages should be jointly trained in a single lexical normalization model. Using a fixed-capacity character-level model across twelve languages, it finds that accuracy peaks when a language is trained with only a few others—typically one to four—and then declines sharply as more languages are added, dropping about forty percent. A control experiment keeping total training data constant shows the decline is due to competition for model capacity rather than data scarcity, and no reliable typological rule predicts the optimal number of co‑training languages.
The study investigates how the choice of tokenizer influences multilingual language models across 54 tokenizers and 123 models. It finds that tokenizer impact is greater for languages with less training data, and that excluding a language from tokenizer training consistently worsens its performance. While reallocating tokenizer training data to lower-resource languages can help, it does not guarantee improvement, and the best tokenizer properties vary by language, enabling predictive screening of tokenizer candidates.
The study evaluates phonologically informed tokenization for German speech recognition using a wav2vec 2.0 backbone fine‑tuned with CTC. Three tokenizer families—pretrained multilingual characters, data‑driven BPE, and phonologically informed units from syllabification and grapheme‑to‑phoneme conversion—are compared across 40 fine‑tunes on three German test sets (in‑domain read, dialectal spontaneous, and standard spontaneous speech). Results show that phonologically informed tokenizers match BPE and character baselines in‑domain, but under domain shift their performance depends on vocabulary size rather than linguistic variation, with syllable‑aware tokenization benefiting dialectal and spontaneous speech at small vocabularies; a phoneme‑level confusion analysis indicates that the acoustic encoder dominates error patterns.
The paper evaluates four metrics—CKA, ANC, GMM dominance per token, and ILO—used to measure cross‑lingual representation sharing in multilingual language models. Across 21 models ranging from 125 M to 14 B parameters, the metrics disagree, and the authors attribute this to anisotropy, where representations cluster in a narrow embedding cone. Only ILO shows a strong, robust correlation with cross‑lingual transfer performance (Spearman’s ρ = 0.90) after controlling for model size, family, and task variation, leading the authors to recommend ILO as the primary metric alongside anisotropy diagnostics.