# Achieving Tokenizer Flexibility in Language Models through Heuristic Adaptation and Supertoken Learning

arXiv: https://arxiv.org/abs/2505.09738 · HTML: https://mlsanity.com/paper/2505.09738/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 14 of its 20 references that are in the map.
- First posted on arXiv 14 May 2025.

## Builds on (most cited first, 14 of 14)

- [Neural Machine Translation of Rare Words with Subword Units](https://mlsanity.com/paper/1508.07909/) (2015)
- [Don't Stop Pretraining: Adapt Language Models to Domains and Tasks](https://mlsanity.com/paper/2004.10964/) (2020)
- [Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus](https://mlsanity.com/paper/2104.08758/) (2021)
- [MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers](https://mlsanity.com/paper/2305.07185/) (2023)
- [WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models](https://mlsanity.com/paper/2112.06598/) (2021)
- [SuperBPE: Space Travel for Language Models](https://mlsanity.com/paper/2503.13423/) (2025)
- [As Good as New. How to Successfully Recycle English GPT-2 to Make Models for Other Languages](https://mlsanity.com/paper/2012.05628/) (2020)
- [Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of LLMs for Low-Resource NLP](https://mlsanity.com/paper/2408.04303/) (2024)
- [Tamil-Llama: A New Tamil Language Model Based on Llama 2](https://mlsanity.com/paper/2311.05845/) (2023)
- [Zero-Shot Tokenizer Transfer](https://mlsanity.com/paper/2405.07883/) (2024)
- [Efficient Language Model Training through Cross-Lingual and Progressive Transfer Learning](https://mlsanity.com/paper/2301.09626/) (2023)
- [An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference](https://mlsanity.com/paper/2402.10712/) (2024)
- [LlamaTurk: Adapting Open-Source Generative Large Language Models for Low-Resource Language](https://mlsanity.com/paper/2405.07745/) (2024)
- [ReTok: Replacing Tokenizer to Enhance Representation Efficiency in Large Language Model](https://mlsanity.com/paper/2410.04335/) (2024)

## Built on by (most cited first, 4 of 4)

- [One Tokenizer To Rule Them All: Emergent Language Plasticity via Multilingual Tokenizers](https://mlsanity.com/paper/2506.10766/) (2025)
- [The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems](https://mlsanity.com/paper/2602.15382/) (2026)
- [In-Place Tokenizer Expansion for Pre-trained LLMs](https://mlsanity.com/paper/2607.15232/) (2026)
- [Writing-System-Level Tokenizer Adaptation for Byte-Level BPE](https://mlsanity.com/paper/2608.00582/) (2026)

## Abstract (arXiv)

Pretrained language models (LLMs) are often constrained by their fixed tokenization schemes, leading to inefficiencies and performance limitations, particularly for multilingual or specialized applications. This tokenizer lock-in presents significant challenges. standard methods to overcome this often require prohibitive computational resources. Although tokenizer replacement with heuristic initialization aims to reduce this burden, existing methods often require exhaustive residual fine-tuning and still may not fully preserve semantic nuances or adequately address the underlying compression inefficiencies. Our framework introduces two innovations: first, Tokenadapt, a model-agnostic tokenizer transplantation method, and second, novel pre-tokenization learning for multi-word Supertokens to enhance compression and reduce fragmentation. Tokenadapt initializes new unique token embeddings via a hybrid heuristic that combines two methods: a local estimate based on subword decomposition using the old tokenizer, and a global estimate utilizing the top-k semantically similar tokens from the original vocabulary. This methodology aims to preserve semantics while significantly minimizing retraining requirements. Empirical investigations validate both contributions: the transplantation heuristic successfully initializes unique tokens, markedly outperforming conventional baselines and sophisticated methods including Transtokenizer and ReTok, while our Supertokens achieve notable compression gains. Our zero-shot perplexity results demonstrate that the TokenAdapt hybrid initialization consistently yields lower perplexity ratios compared to both ReTok and TransTokenizer baselines across different base models and newly trained target tokenizers. TokenAdapt typically reduced the overall perplexity ratio significantly compared to ReTok, yielding at least a 2-fold improvement in these aggregate scores.