Tokenization: BPE, WordPiece, and the LLM era

Summary

Tokenization splits text into discrete units (tokens). Modern LLMs use subword schemes (BPE, WordPiece, SentencePiece) that decompose rare words into subword pieces while preserving common words as single tokens.

Tokenization defines the model’s input and output space, it’s irreversibly baked into a trained model.
Token count drives both cost and latency at inference. Different tokenizers give different counts for the same text (sometimes by 30%+).
Tokenization mismatches between training and serving cause silent quality regressions that are hard to debug.
Production LLM bugs in multilingual quality, code performance, and unusual-character handling often involve tokenization.

The main families

BPE (Byte-Pair Encoding)

Original from text compression (1994), adapted for NLP. Algorithm:

Start with a vocabulary of single characters.
Find the most frequent adjacent pair in the corpus.
Merge that pair into a new vocabulary item.
Repeat until target vocabulary size.

Learning objective: trace how corpus-wide pair counts create ranked, reusable BPE tokens that compress common words while leaving rarer words as multiple pieces.

Learning objective · toy byte-level corpus

How does one frequent byte pair grow into reusable subword and whole-word tokens?

START · COUNT ADJACENT PAIRS ACROSS ALL OCCURRENCES

Corpus: m a p × 5 · m a p s × 3 · c a p × 2

Largest count: a + p = 5 + 3 + 2 = 10
Rank 1 merge: add ap to the vocabulary.

ROUND 2 · RECOUNT AFTER APPLYING `a + p → ap`

Corpus now: m ap × 5 · m ap s × 3 · c ap × 2

Largest count: m + ap = 5 + 3 = 8
Rank 2 merge: add map to the vocabulary.

ROUND 3 · RECOUNT AFTER APPLYING `m + ap → map`

Corpus now: map × 5 · map s × 3 · c ap × 2

Largest count: map + s = 3
Rank 3 merge: add maps to the vocabulary.

APPLY · REPLAY MERGES IN LEARNED RANK ORDER

maps → [maps] · common whole word
cap → [c] [ap] · shared learned ending
traps → [t] [r] [ap] [s] · rarer word, more pieces

Read it this way: read downward. At each round, count pairs in the current segmentation, merge the largest count everywhere, then recount. The learned order matters at encoding time: ap forms before map, and map before maps. Frequent patterns therefore become larger tokens; an unfamiliar word still falls back to available byte pieces. Original worked example informed by Sennrich et al. (2016) and the tiktoken documentation.

Result: common words become single tokens; rare words decompose into subwords; arbitrary text is always representable (down to characters).

Used in: GPT-2/3/4, LLaMA family, most modern LLMs.

Byte-level BPE

BPE applied at the byte level, not the character level. Vocabulary starts with 256 byte values; merges happen on byte sequences. The advantage: any UTF-8 text is representable without unknown tokens, and the base vocabulary is fixed.

Used in: GPT-2 onward, RoBERTa, most modern LLMs. This is what “BPE” almost always means in 2026.

WordPiece

Similar to BPE but uses a likelihood-based merging criterion (maximize the probability of the corpus given the vocabulary) instead of frequency. Tokens not at word start are prefixed with ##.

Used in: BERT, distilBERT.

SentencePiece (Unigram)

A different paradigm: starts with a large vocabulary, removes pieces that contribute least to corpus likelihood. Trains on raw text with no pre-tokenization (handles whitespace and punctuation as part of the algorithm). Implementation by Google, often paired with their models.

Used in: T5, mT5, ALBERT, multilingual models.

Tiktoken / cl100k / o200k

Tiktoken is OpenAI’s BPE implementation. Their tokenizers (cl100k_base, o200k_base) are byte-level BPE optimized for code and multilingual text. Available open-source.

What an interviewer expects you to say

If asked about tokenization:

Define subword tokenization and why it solves the OOV (out-of-vocabulary) problem.
Explain BPE algorithm (frequency-based merging).
Distinguish character-level vs byte-level BPE.
Mention vocab size trade-offs (small vocab → more tokens per word; large vocab → more parameters in embedding/output layer).
Mention that tokenization choice is fixed at training time: you can’t change it post-training without retraining.

Bonus depth: mention the multilingual tokenization issue (English-trained tokenizers have very different per-character costs for non-Latin scripts, sometimes 5-10x more tokens for the same content).

Common confusions

“Tokens are words.” No. Tokens are arbitrary pieces. “tokenization” itself might be one or two tokens depending on the tokenizer.
“Switching tokenizer = small change.” It’s a categorical change. The model has to be retrained.
“All tokenizers are roughly equivalent in cost.” No. cl100k vs LLaMA tokenizer can give 10-30% different token counts on the same text. For multilingual and code-heavy content, the gap is much larger.
“BPE is the same as WordPiece.” Similar idea, different merge criterion (frequency vs likelihood). The differences in practice are small.

The LLM-era tokenization gotchas

A few tokenization issues that bite in production:

Tokenizer drift between training and serving. Different versions of the same tokenizer family can subtly differ. Always pin the exact tokenizer version with the model.
Inconsistent BOS/EOS tokens. Different chat-tuned models expect different special tokens for system prompts, user turns, assistant turns. Send the wrong format → quality regression. Use the model’s chat template, not your own.
Multilingual cost asymmetry. English text is ~4 chars/token; Japanese can be ~1 char/token; some scripts are even worse. A 1000-character Japanese document can cost 10x more than a 1000-character English document.
Code tokenization. Code-aware tokenizers handle indentation, common keywords, and operators efficiently. Non-code tokenizers (older BERT, etc.) treat code very inefficiently.
Unicode normalization. Different unicode normalizations of the “same” string (composed vs decomposed) can produce different token sequences. NFC normalization is standard but worth verifying for your domain.
Number tokenization. Some tokenizers split numbers digit-by-digit (great for arithmetic), others treat numbers as multi-digit tokens (worse for arithmetic). LLaMA and modern models tend to split digits.

Why interviewers ask

Tokenization questions test:

Whether you know the algorithm vs just the vocabulary file.
Whether you’ve debugged a real tokenization bug in production.
Whether you know modern LLM tokenization choices (byte-level BPE, digit splitting, code-aware vocab).
Whether you’re aware of multilingual fairness/cost issues.

A common follow-up asks how to handle an inefficient tokenizer for medical, legal, or code text. Extend the tokenizer and continue pretraining, or train a new tokenizer on domain text and retrain the model.