TokLens: A Multilingual Lens on Tokenizer Quality for LLMs

Accepted to ACL 2026 SRW. 🎉

Open-source toolkit for evaluating tokenizer quality across languages using six intrinsic metrics. We evaluate 24 tokenizers from major LLM families across 15 typologically diverse languages and correlate with downstream performance.

Key Findings

1. Stark cross-lingual disparities persist. GPT-2 produces 56x more tokens per word in Japanese than English. Qwen2.5 and Gemma-2 reduce this gap to under 4x.
2. No metric predicts English benchmark performance after controlling for model size (Bonferroni-corrected). Tokenizer quality does not drive English leaderboard scores.
3. STRR significantly predicts multilingual performance. On MMLU-ProX, linear mixed-effects models show STRR has a large positive effect (β = +5.7, z = 18.5, p < 0.001).
4. Higher STRR correlates with steeper scaling. A controlled experiment on the Qwen2.5 family (fixed tokenizer, varying model size) shows languages with higher STRR scale more steeply (ρ = 0.91, p < 0.001).

Benchmark correlations

Per-language scaling slope vs. tokenizer metrics (Qwen2.5 family)

Metrics

Models and Languages

22 models with Open LLM Leaderboard v2 scores, plus 2 extra tokenizers (Qwen3, DeepSeek-V3) for metric-only analysis.

15 languages across 6 scripts: English, Chinese, Japanese, Arabic, Hindi, German, Turkish, Korean, Thai, Russian, French, Spanish, Portuguese, Vietnamese, Indonesian.

Quickstart

pip install toklens

from toklens import Analyzer

analyzer = Analyzer.from_pretrained("meta-llama/Llama-3.1-8B")
report = analyzer.evaluate(langs=["en", "zh", "ja", "ar"])
report.print_table()

toklens eval meta-llama/Llama-3.1-8B --langs en zh ja ar
toklens compare meta-llama/Llama-2-7b-hf meta-llama/Llama-3.1-8B

Experiments

Reproduces the full evaluation. Run steps in order:

uv run python -m experiments.pipeline.01_collect_benchmarks
uv run python -m experiments.pipeline.02_compute_metrics
uv run python -m experiments.pipeline.03_correlation
uv run python -m experiments.pipeline.04_figures

Supplementary analyses (LME models, Qtok comparison, BPB, Qwen scaling) are in experiments/analyses/. See experiments/README.md for details.

License

MIT