Mehdi Ali, Michael Fromm, Klaudia Thellmann, Richard Rutmann, Max Lübbering, Johannes Leveling, Katrin Klug, Jan Ebert, Niclas Doll, Jasper Buschhoff, Charvi Jain, Alexander Weber, Lena Jurkschat, Hammam Abdelwahab, Chelsea John, Pedro Ortiz Suarez, Malte Ostendorff, Samuel Weinbach, Rafet Sifa, Stefan Kesselheim, Nicolas Flores-Herr
| Challenge: | Recent success of large language models has been driven by curating the training dataset composition, scaling of model architectures and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. |
| Approach: | They conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale. |
| Outcome: | The proposed model can significantly impact the model's downstream performance and training costs. |
Similar Papers
How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models (2021.acl-long)
Copied to clipboard
| Challenge: | Using pretraining data, we find that a designated monolingual tokenizer plays an equally important role in the downstream performance of the model. |
| Approach: | They propose to compare pretrained multilingual models with their monolingual counterparts on a set of five diverse monolingual downstream tasks. |
| Outcome: | The proposed models offer previously unmatched performance in all NLP tasks. |
Tokenization is Sensitive to Language Variation (2025.findings-acl)
Copied to clipboard
| Challenge: | Variation in language is often linked to regional, social, and contextual factors. |
| Approach: | They propose a method to estimate tokenizer impact on downstream LLM performance . they pre-train BERT models with the popular Byte-Pair Encoding algorithm . |
| Outcome: | The proposed model improves on Rényi efficiency and other metrics on language variation. |
One Tokenizer To Rule Them All: Emergent Language Plasticity via Multilingual Tokenizers (2026.acl-long)
Copied to clipboard
Diana Abagyan, Alejandro R. Salamanca, Andres Felipe Cruz-Salinas, Kris Cao, Hangyu Lin, Acyr Locatelli, Marzieh Fadaee, Ahmet Üstün, Sara Hooker
| Challenge: | Existing approaches to train multilingual large language models for many languages at once are limited due to limited model capacity, scarce high-quality data, and compute constraints. |
| Approach: | They propose to use a universal tokenizer to improve language plasticity and adaptability to new languages by up to 20%. |
| Outcome: | The proposed tokenizer improves language plasticity and improves plasticity towards languages that are completely unseen in the tokenizer and pretraining, by up to 5% win rate gain. |
Beyond Text Compression: Evaluating Tokenizers Across Scales (2025.acl-long)
Copied to clipboard
| Challenge: | Language models rely on tokenizers to convert text into machine-interpretable tokens, which shape the statistical patterns that language models learn to estimate. |
| Approach: | They propose to use Zipf's law to measure tokenizer performance by combining several metrics to capture multiple aspects of tokenizer behavior. |
| Outcome: | The proposed metrics correlate more strongly with downstream performance than text compression when modeling unseen languages. |
Adaptation Odyssey in LLMs: Why Does Additional Pretraining Sometimes Fail to Improve? (2024.emnlp-main)
Copied to clipboard
| Challenge: | In the last decade, the generalization and adaptation abilities of deep learning models were evaluated on fixed training and test distributions. |
| Approach: | They propose to train large language models on unlabeled text corpora and train them online. |
| Outcome: | The proposed model training on a text domain could degrade its perplexity on the test portion of the same domain. |
Not-Just-Scaling Laws: Towards a Better Understanding of the Downstream Impact of Language Model Design Decisions (2025.emnlp-main)
Copied to clipboard
Emmy Liu, Amanda Bertsch, Lintang Sutawika, Lindia Tjuatja, Patrick Fernandes, Lara Marinov, Michael Chen, Shreya Singhal, Carolin Lawrence, Aditi Raghunathan, Kiril Gashteovski, Graham Neubig
| Challenge: | Language model performance is largely dependent on pretraining decisions, but scaling laws based on only these two aspects do not always explain downstream task performance. |
| Approach: | They meta-analyze 92 open-source pretrained models to quantify their impact on performance. |
| Outcome: | The framework lays a foundation for more systematic investigation of how model development choices shape final capabilities. |
Tokenization Impacts Multilingual Language Modeling: Assessing Vocabulary Allocation and Overlap Across Languages (2023.findings-acl)
Copied to clipboard
| Challenge: | Multilingual language models perform surprisingly well in a variety of NLP tasks for diverse languages. |
| Approach: | They propose to evaluate the quality of lexical representation and vocabulary overlap observed in sub-word tokenizers. |
| Outcome: | The proposed criteria show that the overlap of vocabulary across languages can be detrimental to certain downstream tasks. |
Exploring Design Choices for Building Language-Specific LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Prior work focused on building multilingual models that cover a broad spectrum of languages. |
| Approach: | They conduct systematic experiments on how design choices impact the adapted LLM, both in terms of efficiency and end task performance. |
| Outcome: | The proposed model performs better on English-centric models than multilingual models despite poor performance on low-resource languages. |
A Multi-dimensional Evaluation of Tokenizer-free Multilingual Pretrained Models (2023.findings-eacl)
Copied to clipboard
| Challenge: | Recent work on tokenizer-free models shows promising results in cross-lingual transfer . previous work focused on reporting accuracy on a limited set of tasks and data settings . |
| Approach: | They compare tokenizer-free and subword-based models using various dimensions . they find subword models are still the most practical choice in many settings . |
| Outcome: | The proposed model improves cross-lingual transfer and reduces engineering overhead. |
Token Weighting for Long-Range Language Modeling (2025.findings-naacl)
Copied to clipboard
| Challenge: | Many applications of large language models (LLMs) require long-context understanding, but models still struggle with such tasks. |
| Approach: | They propose token-weighting schemes that assign different weights to each training token in the loss, generalizing existing works. |
| Outcome: | The proposed methods compare confidences of a long-context and short-concept model and show that non-uniform loss weights improve the long-constability of LLMs. |