Large language models are increasingly deployed as general-purpose educational and technical assistance systems, but their underlying infrastructure does not treat languages equally. One underexamined source of disparity is tokenization: semantically equivalent content can require substantially different token counts across languages, affecting API cost, latency, and usable context length before a model is invoked.
Large language models (LLMs) process text through subword tokenizers rather than directly reading characters or words. Because these tokenizers are trained predominantly on English-centric corpora, they introduce a systematic and often overlooked disadvantage for many non-English languages.
arXiv:2607. 24276v1 Announce Type: cross Abstract: Large language models (LLMs) process text through subword tokenizers rather than directly reading characters or words.
By Priyansh Srivastava
arXiv:2607. 23319v1 Announce Type: cross Abstract: Standard subword tokenization algorithms such as Byte-Pair Encoding (BPE) and SentencePiece are trained predominantly on modern language corpora and produce inefficient segmentations when applied to classical Indian languages.
By Poornima Kumaresan, Pavithra Muruganantham, Lakshmi Rajendran, Santhosh Sivasubramani
The paper shows that HuggingFace’s ByteLevel pre‑tokenizer, which treats a word as a sequence of Unicode letters, splits abugida scripts at every vowel sign, creating a training‑free lower bound on tokenizer fertility. Across 26 languages, all 17 abugidas exhibit increased token counts (up to 9×), while Latin, Cyrillic, Hangul, and Han remain unchanged. The authors demonstrate that correcting the character class reduces Nepali token counts, improves model performance, and that this issue is widespread in popular HuggingFace models.
By Sajal Regmi, Siddhartha Pudasaini, Chetan Phakami Pun
arXiv:2608.30092v1 Announce Type: cross
Abstract: We present Arkios, a 1.04B-parameter dense transformer pretrained from scratch on 150B tokens of bilingual English-Nepali text, using a custom single...
By Sajal Regmi, Siddhartha Pudasaini, Chetan Phakami Pun
arXiv:2608.16344v3 Announce Type: replace
Abstract: Indic quality estimation (QE) and automatic post-editing (APE) data is spread across separate releases, so no single resource supports training and...
By Diptesh Kanojia, Archchana Sindhujan, Sourabh Deoghare, Daria Sokova, Shenbin Qian, Girish Koushik, Tharindu Ranasinghe, Constantin Or\u{a}san, Chrysoula Zerva, Ricardo Rei, Fr\'ed\'eric Blain, Andr\'e F. T. Martins, Marco Turchi, Matteo Negri, Anoop Kunchukuttan, Mitesh M. Khapra, Pushpak Bhattacharyya
Indic quality estimation (QE) and automatic post-editing (APE) data is spread across separate releases, so no single resource supports training and evaluation across tasks and language pairs on one footing. We consolidate the WMT 2020--2024 shared-task lineage with an extended English--Malayalam resource into \indicqe: $126{,}754$ instances over nine directional pairs, with up to four label types aligned on the same segment, a direct assessment, a human post-edit, word-level OK/BAD tags and an error explanation, and a test set stratified over four difficulty axes.
IndicQE-APE is a consolidated benchmark that unifies quality estimation (QE) and automatic post‑editing (APE) data for nine Indic language pairs, comprising 126,754 instances with multiple aligned labels such as direct assessment, human post‑edit, word‑level OK/BAD tags, and error explanations. The dataset includes a stratified test set across four difficulty axes and supports training and evaluation of six prompted large language models, three COMET metrics, and three APE systems. Experiments reveal that segments with conflicting holistic and token‑level quality signals are consistently ranked lower, while annotator disagreement shows no effect when controlled for score distribution.
whyItMatters":"The benchmark provides a unified resource for training and evaluating QE and APE across Indic languages, enabling consistent comparison of models and metrics on a shared dataset."
By Diptesh Kanojia, Archchana Sindhujan, Sourabh Deoghare, Daria Sokova, Shenbin Qian, Girish Koushik, Tharindu Ranasinghe, Constantin Or\u{a}san, Chrysoula Zerva, Ricardo Rei, Fr\'ed\'eric Blain, Andr\'e F. T. Martins, Marco Turchi, Matteo Negri, Anoop Kunchukuttan, Mitesh M. Khapra, Pushpak Bhattacharyya
arXiv:2609.12960v1 Announce Type: new
Abstract: Sanskrit fuses case, number, person and tense into word endings and chains clauses into compounds, so it is information-dense per word. Whether that de...
By Devansh Sharma
arXiv:2608.21384v1 Announce Type: cross
Abstract: Modern multilingual tokenizers often fragment Ukrainian and other underrepresented Cyrillic-script languages more heavily than English, creating disp...
By Ivan Dobrovolskyi
arXiv:2606. 24460v1 Announce Type: cross Abstract: Commercial large language models bill, scale latency, and budget context per token.
By Olaoye Anthony Somide