# HashFormers: Towards Vocabulary-independent Pre-trained Transformers

arXiv: https://arxiv.org/abs/2210.07904 · HTML: https://mlsanity.com/paper/2210.07904/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- Builds on 28 of its 39 references that are in the map.
- First posted on arXiv 14 Oct 2022.

## Builds on (most cited first, 28 of 28)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding](https://mlsanity.com/paper/1804.07461/) (2018)
- [Neural Machine Translation of Rare Words with Subword Units](https://mlsanity.com/paper/1508.07909/) (2015)
- [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://mlsanity.com/paper/1906.08237/) (2019)
- [SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing](https://mlsanity.com/paper/1808.06226/) (2018)
- [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://mlsanity.com/paper/1901.02860/) (2019)
- [Reformer: The Efficient Transformer](https://mlsanity.com/paper/2001.04451/) (2020)
- [Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation](https://mlsanity.com/paper/1609.08144/) (2016)
- [Energy and Policy Considerations for Deep Learning in NLP](https://mlsanity.com/paper/1906.02243/) (2019)
- [Aligning Books and Movies: Towards Story-like Visual Explanations by Watching Movies and Reading Books](https://mlsanity.com/paper/1506.06724/) (2015)
- [Green AI](https://mlsanity.com/paper/1907.10597/) (2019)
- [Datasets: A Community Library for Natural Language Processing](https://mlsanity.com/paper/2109.02846/) (2021)
- [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://mlsanity.com/paper/2105.13626/) (2021)
- [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://mlsanity.com/paper/2103.06874/) (2021)
- [Combating Adversarial Misspellings with Robust Word Recognition](https://mlsanity.com/paper/1905.11268/) (2019)
- [Charformer: Fast Character Transformers via Gradient-based Subword Tokenization](https://mlsanity.com/paper/2106.12672/) (2021)
- [Rethinking embedding coupling in pre-trained language models](https://mlsanity.com/paper/2010.12821/) (2020)
- [Compressing Large-Scale Transformer-Based Models: A Case Study on BERT](https://mlsanity.com/paper/2002.11985/) (2020)
- [CharacterBERT: Reconciling ELMo and BERT for Word-Level Open-Vocabulary Representations From Characters](https://mlsanity.com/paper/2010.10392/) (2020)
- [Adv-BERT: BERT is not robust on misspellings! Generating nature adversarial samples on BERT](https://mlsanity.com/paper/2003.04985/) (2020)
- [Hash Embeddings for Efficient Word Representations](https://mlsanity.com/paper/1709.03933/) (2017)
- [Compressing Word Embeddings via Deep Compositional Code Learning](https://mlsanity.com/paper/1711.01068/) (2017)
- [Frustratingly Simple Pretraining Alternatives to Masked Language Modeling](https://mlsanity.com/paper/2109.01819/) (2021)
- [How does the pre-training objective affect what large language models learn about linguistic properties?](https://mlsanity.com/paper/2203.10415/) (2022)
- [ProFormer: Towards On-Device LSH Projection Based Transformers](https://mlsanity.com/paper/2004.05801/) (2020)
- [Compressing Transformer-Based Semantic Parsing Models using Compositional Code Embeddings](https://mlsanity.com/paper/2010.05002/) (2020)

## Built on by (most cited first, 6 of 6)

- [T-FREE: Subword Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings](https://mlsanity.com/paper/2406.19223/) (2024)
- [Vocabulary-level Memory Efficiency for Language Model Fine-tuning](https://mlsanity.com/paper/2309.08708/) (2023)
- [Understanding the Role of Input Token Characters in Language Models: How Does Information Loss Affect Performance?](https://mlsanity.com/paper/2310.17271/) (2023)
- [Deconstructing Attention: Investigating Design Principles for Effective Language Modeling](https://mlsanity.com/paper/2510.11602/) (2025)
- [Pit One Against Many: Leveraging Attention-head Embeddings for Parameter-efficient Multi-head Attention](https://mlsanity.com/paper/2310.07911/) (2023)
- [MultiHashFormer: Hash-based Generative Language Models](https://mlsanity.com/paper/2606.28057/) (2026)

## Abstract (arXiv)

Transformer-based pre-trained language models are vocabulary-dependent, mapping by default each token to its corresponding embedding. This one-to-one mapping results into embedding matrices that occupy a lot of memory (i.e. millions of parameters) and grow linearly with the size of the vocabulary. Previous work on on-device transformers dynamically generate token embeddings on-the-fly without embedding matrices using locality-sensitive hashing over morphological information. These embeddings are subsequently fed into transformer layers for text classification. However, these methods are not pre-trained. Inspired by this line of work, we propose HashFormers, a new family of vocabulary-independent pre-trained transformers that support an unlimited vocabulary (i.e. all possible tokens in a corpus) given a substantially smaller fixed-sized embedding matrix. We achieve this by first introducing computationally cheap hashing functions that bucket together individual tokens to embeddings. We also propose three variants that do not require an embedding matrix at all, further reducing the memory requirements. We empirically demonstrate that HashFormers are more memory efficient compared to standard pre-trained transformers while achieving comparable predictive performance when fine-tuned on multiple text classification tasks. For example, our most efficient HashFormer variant has a negligible performance degradation (0.4% on GLUE) using only 99.1K parameters for representing the embeddings compared to 12.3-38M parameters of state-of-the-art models.