2026· Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM 2026)· pp. 221-228· 0 citations· 16 references
TL;DR
This paper presents a multilingual evaluation of parity-aware to-kenization under worst-N optimization, extending PA-BPE to jointly optimize over the N worst-compressed languages.
Abstract
Improving the fairness of a language model is a goal that applies at every level of the model. In this paper, we evaluate a method targeting a foundational level: tokenization. We present a multilingual evaluation of parity-aware to-kenization under worst-N optimization, extending PA-BPE to jointly optimize over the N worst-compressed languages. We evaluate this formulation for N > 1 across vocabulary sizes of 16K and 32K on the languages from the flores+ benchmark, using metrics that capture both efficiency and structural alignment. Our results reveal that the effects of increasing N are inconsistent across metrics and do not lead to major gains. Efficiency-oriented and boundary-level metrics show a modest tendency to improve at higher values of N , while structural alignment metrics (such as AST alignment and boundary crossing) exhibit no clear pattern, suggesting that compression fairness and linguistic structure are mainly orthogonal objectives. Script-level analysis further reveals uneven effects across writing systems, with several non-Latin scripts showing greater sensitivity to increasing N .
Language model tokenizers are typically selected with minimal evaluation, despite the fact that their design choices directly impact model capabilities. This can be partly attributed to a limited understanding of which tokenizer properties affect which aspects of downstream performance. We introduce TokEval, a framework of tokenizer evaluation metrics that goes beyond standard measures like fertility and compression rate to capture linguistically and structurally meaningful properties, e.g., UTF-8 character boundary integrity and digit place-value boundary alignment for mathematics. To validate whether these metrics are predictive of downstream model performance, we conduct controlled language model pretraining experiments, varying solely the tokenizers'training data mixture, pretokenization strategy, and training algorithm. We evaluate the resulting models on bits-per-byte (a tokenizer-agnostic version of perplexity) and several benchmarks, spanning linguistic understanding, mathematical reasoning, and code generation. Our experiments suggest that different intrinsic properties have different impacts on model abilities: information-theoretic metrics predict language modeling abilities (Spearman rho up to 0.80), while structure-sensitive metrics, such as those measuring digit and line-break handling, correlate with task accuracy. We hope TokEval enables more principled tokenizer evaluation, replacing pretraining sweeps with intrinsic measurement wherever the two agree.
It is shown that English-only safety evaluations are insufficient; they require accounting for script family, perturbation type, and per-language alignment coverage, and a geometric mechanistic analysis of refusal failure across language tiers.
Large language models (LLMs) have achieved remarkable success in high-resource languages, yet their performance on Traditional Mongolian remains highly limited. A primary bottleneck is the absence of a systematic evaluation framework, which precludes quantitative comparison and obscures directions for model optimization. In this paper, we introduce TM-Bench, the first comprehensive benchmark for LLMs on Traditional Mongolian. TM-Bench adopts a hybrid construction strategy consisting of human-verified Translation-based Adaptation, Expert-Original Authoring, and Semi-automated Synthesis. It comprises 18,357 instances spanning five tasks across both natural language understanding and generation to evaluate models' reasoning, knowledge application, and linguistic proficiency. We conduct systematic evaluations across representative model families. The results show that on understanding tasks, model performance lags significantly behind high-resource languages, with only a few models performing slightly above the random baseline. For generation tasks, both automatic metrics and double-blind human evaluations reveal severe semantic collapse, failing to generate coherent text and often producing unreadable gibberish. These findings underscore the critical role of TM-Bench as a foundational infrastructure for evaluating LLMs in Traditional Mongolian and catalyzing future model optimization. Our benchmark and code are available at https://github.com/gao1948083886/TM-Bench.
Zhenjie Gao, Feilong Bao, Aruukhan Bai et al.· Annual International ACM SIG...· 0 citations
This work addresses the challenge of building robust foundation models for underrepresented Central Asian languages by presenting GigaAM Multilingual, a Conformer encoder pre-trained on 2M hours of audio using a HuBERT-style objective, and introduces a cluster-level data balancing strategy during pre-training and a domain-aware sampling method during fine-tuning to mitigate head-language dominance.
Andrei Kuzmenko, A. Maximenko, Aleksandr Kutsakov et al.· 0 citations
Aggressive quantization disproportionately harms multilingual capability: in the sub-4B INT3 GPTQ regime, we measure 2-4x larger perplexity degradation on non-English languages than on English. We propose Language-Conditional Dequantization (LCD), a post-hoc method that attaches per-language rank-2 LoRA corrections to the linear layers of an already-quantized model, adding 0.12% parameters per language and training in under 20 minutes on a single GPU. Across Qwen2.5-3B and Llama-3.2-3B, LCD recovers 70-83% of the perplexity gap for non-Latin script languages and 17-28% of the GlobalMMLU accuracy gap, outperforming a language-agnostic correction of equal capacity by 3-9 points on typologically distant languages and a data-free low-rank baseline (LQER) by an order of magnitude. We further identify a perplexity-accuracy disconnect and trace it to where quantization concentrates damage: early-depth errors (Llama) propagate downstream and resist local correction, while late-depth errors (Qwen) do not. A layer-restricted variant of LCD validates this mechanism directly.
MGAL is the first multilingual, granularity- and position-aware long-context benchmark, constructed from United Nations reports spanning 8K to 128K tokens across the six official UN languages, and finds that LLMs perform well at word-level tasks but struggle with coarser-grained ones.
Chunhan Li, Chenglin Xu, Zongyang Zhang et al.· 0 citations
