Defragmenting Language Models: An Interpretability-based Approach for Vocabulary Expansion

• URL: http://arxiv.org/abs/2604.16656v1
• Date: Fri, 17 Apr 2026 19:21:11 GMT
• Title: Defragmenting Language Models: An Interpretability-based Approach for Vocabulary Expansion
• Abstract summary: We study interpretability-based vocabulary expansion, a new research direction.<n>We show that interpretability-based methods offer a superior performance-token efficiency trade-off.<n>We propose FragMend to push the efficiency ceiling of interpretability-based expansion.
• Score: 25.257834614765372
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: All languages are equal; when it comes to tokenization, some are more equal than others. Tokens are the hidden currency that dictate the cost and latency of access to contemporary LLMs. However, many languages written in non-Latin scripts observe a poor exchange rate: LLMs take several multiples of tokens to encode the same information in many languages as they do for English. Our analysis reveals that this issue, known as 'token over-fragmentation', persists in modern open-weight LLMs. The standard remedy is vocabulary expansion that adds target language items missing from the model's vocabulary. In this work, we comprehensively study and advance interpretability-based vocabulary expansion, a new research direction. We focus on two core decisions in the vocabulary expansion process: What items should we add? and How should we initialize their corresponding input and output embeddings? First, we question the conventional use of frequency-based methods to choose candidate vocabulary items to add (a decision long treated as settled), and show that interpretability-based methods offer a superior performance-token efficiency trade-off. Next, we strengthen the case for interpretability-based embedding initialization by showing large gains (~20 pts) over baseline initialization methods for several languages written in non-Latin scripts. We identify the phenomenon of "subword detokenization" where models progressively merge fragmented subword tokens into larger subwords across layers. Grounded in our analysis of this phenomenon, we propose FragMend to further push the efficiency ceiling of interpretability-based expansion. We validate the effectiveness of FragMend through comparison against strong baselines and we present extensive analysis of its design choices.

Related papers

• Tokenizing Crosslingual Homographs [8.254230288283258]
  Multilingual language models rely on shared subword vocabularies to represent multiple languages within a limited number of token units.<n>We investigate this limitation through cross-lingual homographs and false friends.<n>We propose a simple tokenizer-level intervention based on language cues: language-specific characters replacing initial characters of shared-vocabulary words.
  arXiv  Detail & Related papers  (2026-07-20T08:37:31Z)
• TokAlign++: Advancing Vocabulary Adaptation via Better Token Alignment [46.819665083185136]
  Inefficient tokenization results in long token-ID sequences and will slow down the training and inference of Large Language Models.<n>We introduce a method named TokAlign++ to improve vocabulary adaptation performance by learning better token alignment lexicon.<n> Experimental results on 15 languages show that our method boosts the multilingual text compression rates and preserves most of the multilingual ability of vanilla models.
  arXiv  Detail & Related papers  (2026-05-13T12:23:24Z)
• False Friends Are Not Foes: Investigating Vocabulary Overlap in Multilingual Language Models [53.01170039144264]
  Subword tokenizers trained on multilingual corpora naturally produce overlapping tokens across languages.<n>Does token overlap facilitate cross-lingual transfer or instead introduce interference between languages?<n>We find that models with overlap outperform models with disjoint vocabularies.
  arXiv  Detail & Related papers  (2025-09-23T07:47:54Z)
• HYPEROFA: Expanding LLM Vocabulary to New Languages via Hypernetwork-Based Embedding Initialization [50.27950279695363]
  Many pre-trained language models (PLMs) exhibit suboptimal performance on mid- and low-resource languages.<n>A common strategy to address this is to introduce new tokens specific to the target languages, initialize their embeddings, and apply continual pre-training on target-language data.<n>We propose HYPEROFA, a hypernetwork-based approach for more adaptive token embedding.
  arXiv  Detail & Related papers  (2025-04-21T19:40:32Z)
• How Can We Effectively Expand the Vocabulary of LLMs with 0.01GB of Target Language Text? [34.57680190813489]
  Large language models (LLMs) have shown remarkable capabilities in many languages beyond English.<n>LLMs require more inference steps when generating non-English text due to their reliance on English-centric tokenizers and vocabulary.<n> Vocabulary expansion with target language tokens is a widely used cross-lingual vocabulary adaptation approach to remedy this issue.
  arXiv  Detail & Related papers  (2024-06-17T12:42:34Z)
• The Ups and Downs of Large Language Model Inference with Vocabulary Trimming by Language Heuristics [74.99898531299148]
  This research examines vocabulary trimming (VT) inspired by restricting embedding entries to the language of interest to bolster time and memory efficiency. We apply two languages to trim the full vocabulary - Unicode-based script filtering and corpus-based selection - to different language families and sizes. It is found that VT reduces the memory usage of small models by nearly 50% and has an upper bound of 25% improvement in generation speed.
  arXiv  Detail & Related papers  (2023-11-16T09:35:50Z)
• Tokenization Impacts Multilingual Language Modeling: Assessing Vocabulary Allocation and Overlap Across Languages [3.716965622352967]
  We propose new criteria to evaluate the quality of lexical representation and vocabulary overlap observed in sub-word tokenizers. Our findings show that the overlap of vocabulary across languages can be actually detrimental to certain downstream tasks.
  arXiv  Detail & Related papers  (2023-05-26T18:06:49Z)
• Translate to Disambiguate: Zero-shot Multilingual Word Sense Disambiguation with Pretrained Language Models [67.19567060894563]
  Pretrained Language Models (PLMs) learn rich cross-lingual knowledge and can be finetuned to perform well on diverse tasks. We present a new study investigating how well PLMs capture cross-lingual word sense with Contextual Word-Level Translation (C-WLT) We find that as the model size increases, PLMs encode more cross-lingual word sense knowledge and better use context to improve WLT performance.
  arXiv  Detail & Related papers  (2023-04-26T19:55:52Z)
• A Vocabulary-Free Multilingual Neural Tokenizer for End-to-End Task Learning [8.052271364177988]
  Subword tokenization is a commonly used input pre-processing step in most recent NLP models. We propose a vocabulary-free neural tokenizer by distilling segmentation information from subword tokenization. Our tokenizer consistently improves performance on multilingual (NLI) and code-switching (sentiment analysis) tasks.
  arXiv  Detail & Related papers  (2022-04-22T16:50:49Z)
• More Than Words: Collocation Tokenization for Latent Dirichlet Allocation Models [71.42030830910227]
  We propose a new metric for measuring the clustering quality in settings where the models differ. We show that topics trained with merged tokens result in topic keys that are clearer, more coherent, and more effective at distinguishing topics than those unmerged models.
  arXiv  Detail & Related papers  (2021-08-24T14:08:19Z)
• Multilingual Autoregressive Entity Linking [49.35994386221958]
  mGENRE is a sequence-to-sequence system for the Multilingual Entity Linking problem. For a mention in a given language, mGENRE predicts the name of the target entity left-to-right, token-by-token. We show the efficacy of our approach through extensive evaluation including experiments on three popular MEL benchmarks.
  arXiv  Detail & Related papers  (2021-03-23T13:25:55Z)
• FILTER: An Enhanced Fusion Method for Cross-lingual Language Understanding [85.29270319872597]
  We propose an enhanced fusion method that takes cross-lingual data as input for XLM finetuning. During inference, the model makes predictions based on the text input in the target language and its translation in the source language. To tackle this issue, we propose an additional KL-divergence self-teaching loss for model training, based on auto-generated soft pseudo-labels for translated text in the target language.
  arXiv  Detail & Related papers  (2020-09-10T22:42:15Z)
• Language-Independent Tokenisation Rivals Language-Specific Tokenisation for Word Similarity Prediction [12.376752724719005]
  Language-independent tokenisation (LIT) methods do not require labelled language resources or lexicons. Language-specific tokenisation (LST) methods have a long and established history, and are developed using carefully created lexicons and training resources. We empirically compare the two approaches using semantic similarity measurement as an evaluation task across a diverse set of languages.
  arXiv  Detail & Related papers  (2020-02-25T16:24:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.