# What do tokens know about their characters and how do they know it?

arXiv: https://arxiv.org/abs/2206.02608 · HTML: https://mlsanity.com/paper/2206.02608/

## Key facts

- Cited by 36 papers in the map (data through 14 Oct 2026).
- Builds on 41 of its 83 references that are in the map.
- First posted on arXiv 6 Jun 2022.

## Builds on (most cited first, 30 of 41)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [Adam: A Method for Stochastic Optimization](https://mlsanity.com/paper/1412.6980/) (2014)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [PyTorch: An Imperative Style, High-Performance Deep Learning Library](https://mlsanity.com/paper/1912.01703/) (2019)
- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [Efficient Estimation of Word Representations in Vector Space](https://mlsanity.com/paper/1301.3781/) (2013)
- [On the Opportunities and Risks of Foundation Models](https://mlsanity.com/paper/2108.07258/) (2021)
- [Neural Machine Translation of Rare Words with Subword Units](https://mlsanity.com/paper/1508.07909/) (2015)
- [The Pile: An 800GB Dataset of Diverse Text for Language Modeling](https://mlsanity.com/paper/2101.00027/) (2020)
- [HuggingFace's Transformers: State-of-the-art Natural Language Processing](https://mlsanity.com/paper/1910.03771/) (2019)
- [SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing](https://mlsanity.com/paper/1808.06226/) (2018)
- [Enriching Word Vectors with Subword Information](https://mlsanity.com/paper/1607.04606/) (2016)
- [LXMERT: Learning Cross-Modality Encoder Representations from Transformers](https://mlsanity.com/paper/1908.07490/) (2019)
- [Multilingual Denoising Pre-training for Neural Machine Translation](https://mlsanity.com/paper/2001.08210/) (2020)
- [Probing Classifiers: Promises, Shortcomings, and Advances](https://mlsanity.com/paper/2102.12452/) (2021)
- [A Primer in BERTology: What we know about how BERT works](https://mlsanity.com/paper/2002.12327/) (2020)
- [Designing and Interpreting Probes with Control Tasks](https://mlsanity.com/paper/1909.03368/) (2019)
- [Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates](https://mlsanity.com/paper/1804.10959/) (2018)
- [End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF](https://mlsanity.com/paper/1603.01354/) (2016)
- [How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models](https://mlsanity.com/paper/2012.15613/) (2020)
- [Analysis Methods in Neural Language Processing: A Survey](https://mlsanity.com/paper/1812.08951/) (2018)
- [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://mlsanity.com/paper/2103.06874/) (2021)
- [Information-Theoretic Probing for Linguistic Structure](https://mlsanity.com/paper/2004.03061/) (2020)
- [BPE-Dropout: Simple and Effective Subword Regularization](https://mlsanity.com/paper/1910.13267/) (2019)
- [Character-Aware Neural Language Models](https://mlsanity.com/paper/1508.06615/) (2015)
- [Between words and characters: A Brief History of Open-Vocabulary Modeling and Tokenization in NLP](https://mlsanity.com/paper/2112.10508/) (2021)
- [Analyzing the Source and Target Contributions to Predictions in Neural Machine Translation](https://mlsanity.com/paper/2010.10907/) (2020)
- [What Kind of Language Is Hard to Language-Model?](https://mlsanity.com/paper/1906.04726/) (2019)
- [You should evaluate your language model on marginal likelihood over tokenisations](https://mlsanity.com/paper/2109.02550/) (2021)
- [Flexible Generation of Natural Language Deductions](https://mlsanity.com/paper/2104.08825/) (2021)

## Built on by (most cited first, 30 of 36)

- [A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders](https://mlsanity.com/paper/2409.14507/) (2024)
- [Character-Aware Models Improve Visual Text Rendering](https://mlsanity.com/paper/2212.10562/) (2022)
- [From Tokens to Words: On the Inner Lexicon of LLMs](https://mlsanity.com/paper/2410.05864/) (2024)
- [Rigorously Assessing Natural Language Explanations of Neurons](https://mlsanity.com/paper/2309.10312/) (2023)
- [Do large language models resemble humans in language use?](https://mlsanity.com/paper/2303.08014/) (2023)
- [CUTE: Measuring LLMs' Understanding of Their Tokens](https://mlsanity.com/paper/2409.15452/) (2024)
- [MrT5: Dynamic Token Merging for Efficient Byte-level Language Models](https://mlsanity.com/paper/2410.20771/) (2024)
- [Representation Engineering for Large-Language Models: Survey and Research Challenges](https://mlsanity.com/paper/2502.17601/) (2025)
- [CharPoet: A Chinese Classical Poetry Generation System Based on Token-free LLM](https://mlsanity.com/paper/2401.03512/) (2024)
- [Broken Tokens? Your Language Model can Secretly Handle Non-Canonical Tokenizations](https://mlsanity.com/paper/2506.19004/) (2025)
- [Inducing Character-level Structure in Subword-based Language Models with Type-level Interchange Intervention Training](https://mlsanity.com/paper/2212.09897/) (2022)
- [What Do Language Models Hear? Probing for Auditory Representations in Language Models](https://mlsanity.com/paper/2402.16998/) (2024)
- [Thinking beyond the anthropomorphic paradigm benefits LLM research](https://mlsanity.com/paper/2502.09192/) (2025)
- [Incorporating Context into Subword Vocabularies](https://mlsanity.com/paper/2210.07095/) (2022)
- [Bolmo: Byteifying the Next Generation of Language Models](https://mlsanity.com/paper/2512.15586/) (2025)
- [Learn Your Tokens: Word-Pooled Tokenization for Language Modeling](https://mlsanity.com/paper/2310.11628/) (2023)
- [StochasTok: Improving Fine-Grained Subword Understanding in LLMs](https://mlsanity.com/paper/2506.01687/) (2025)
- [Interactive Prompt Debugging with Sequence Salience](https://mlsanity.com/paper/2404.07498/) (2024)
- [Knowledge of Pretrained Language Models on Surface Information of Tokens](https://mlsanity.com/paper/2402.09808/) (2024)
- [With Ears to See and Eyes to Hear: Sound Symbolism Experiments with Multimodal Large Language Models](https://mlsanity.com/paper/2409.14917/) (2024)
- [EXECUTE: A Multilingual Benchmark for LLM Token Understanding](https://mlsanity.com/paper/2505.17784/) (2025)
- [Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters](https://mlsanity.com/paper/2506.10641/) (2025)
- [CharED: Character-wise Ensemble Decoding for Large Language Models](https://mlsanity.com/paper/2407.11009/) (2024)
- [Enhancing Character-Level Understanding in LLMs through Token Internal Structure Learning](https://mlsanity.com/paper/2411.17679/) (2024)
- [CharBench: Evaluating the Role of Tokenization in Character-Level Tasks](https://mlsanity.com/paper/2508.02591/) (2025)
- [Fast Byte Latent Transformer](https://mlsanity.com/paper/2605.08044/) (2026)
- [SpeLLM: Character-Level Multi-Head Decoding](https://mlsanity.com/paper/2507.16323/) (2025)
- [Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness](https://mlsanity.com/paper/2603.10771/) (2026)
- [A blind spot for large language models: Supradiegetic linguistic information](https://mlsanity.com/paper/2306.06794/) (2023)
- [Decoding-Free Sampling Strategies for LLM Marginalization](https://mlsanity.com/paper/2510.20208/) (2025)

## Abstract (arXiv)

Pre-trained language models (PLMs) that use subword tokenization schemes can succeed at a variety of language tasks that require character-level information, despite lacking explicit access to the character composition of tokens. Here, studying a range of models (e.g., GPT- J, BERT, RoBERTa, GloVe), we probe what word pieces encode about character-level information by training classifiers to predict the presence or absence of a particular alphabetical character in a token, based on its embedding (e.g., probing whether the model embedding for "cat" encodes that it contains the character "a"). We find that these models robustly encode character-level information and, in general, larger models perform better at the task. We show that these results generalize to characters from non-Latin alphabets (Arabic, Devanagari, and Cyrillic). Then, through a series of experiments and analyses, we investigate the mechanisms through which PLMs acquire English-language character information during training and argue that this knowledge is acquired through multiple phenomena, including a systematic relationship between particular characters and particular parts of speech, as well as natural variability in the tokenization of related strings.