Pith. sign in

REVIEW 11 cited by

Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2402.14903 v1 pith:LH64GGGI submitted 2024-02-22 cs.CL cs.LG

classification cs.CLcs.LG
keywords tokenizationmodelmodelsarithmeticinputleft-to-rightllmsperformance
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Tokenization, the division of input text into input tokens, is an often overlooked aspect of the large language model (LLM) pipeline and could be the source of useful or harmful inductive biases. Historically, LLMs have relied on byte pair encoding, without care to specific input domains. With the increased use of LLMs for reasoning, various number-specific tokenization schemes have been adopted, with popular models like LLaMa and PaLM opting for single-digit tokenization while GPT-3.5 and GPT-4 have separate tokens for each 1-, 2-, and 3-digit numbers. In this work, we study the effect this choice has on numerical reasoning through the use of arithmetic tasks. We consider left-to-right and right-to-left tokenization for GPT-3.5 and -4, finding that right-to-left tokenization (enforced by comma separating numbers at inference time) leads to largely improved performance. Furthermore, we find that model errors when using standard left-to-right tokenization follow stereotyped error patterns, suggesting that model computations are systematic rather than approximate. We show that the model is able to convert between tokenizations easily, thus allowing chain-of-thought-inspired approaches to recover performance on left-to-right tokenized inputs. We also find the gap between tokenization directions decreases when models are scaled, possibly indicating that larger models are better able to override this tokenization-dependent inductive bias. In summary, our work performs the first study of how number tokenization choices lead to differences in model performance on arithmetic tasks, accompanied by a thorough analysis of error patterns. We hope this work inspires practitioners to more carefully ablate number tokenization-related choices when working towards general models of numerical reasoning.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 11 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Beyond In-Context Learning: Aligning Long-form Generation of Large Language Models via Task-Inherent Attribute Guidelines

    cs.CL 2025-06 conditional novelty 7.0 of 10

    LongGuide automatically learns task-specific quality and length guidelines from small training sets, significantly improving LLM long-form generation.

  2. Variable-length Gray codes for the Natural Numbers

    cs.IT 2026-07 accept novelty 6.0 of 10

    Every finite digit string gets exactly one number, and consecutive numbers differ by a single edit — a complete, locally stable variable-length Gray code.

  3. The Effect of Scripts and Formats on LLM Numeracy

    cs.CL 2026-01 conditional novelty 6.0 of 10

    LLM arithmetic accuracy falls sharply when numerals leave the familiar Hindu–Arabic format, and few-shot prompting with examples narrows most of that gap.

  4. Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance

    cs.CL 2025-07 conditional novelty 6.0 of 10

    Falcon-H1 reports competitive benchmark scores for a 0.5B to 34B family of parallel hybrid attention/Mamba-2 models, claiming 2x to 4x parameter efficiency versus dense transformers.

  5. The Decrypto Benchmark for Multi-Agent Reasoning and Theory of Mind

    cs.AI 2025-06 conditional novelty 6.0 of 10

    A new interactive language-game benchmark shows LLMs lag behind simple word-embedding baselines and that newer reasoning models regress on theory-of-mind tasks.

  6. Incorporating Domain Knowledge into Materials Tokenization

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A domain-knowledge-weighted tokenizer, MATTER, preserves material concepts and improves materials NLP performance by 4% on generation and 2% on classification tasks.

  7. BPE Stays on SCRIPT: Structured Encoding for Robust Multilingual Pretokenization

    cs.CL 2025-05 conditional novelty 6.0 of 10

    A Unicode script and category based character encoding with constrained merging achieves compression competitive with byte-level BPE while removing the byte-premium penalty for non-Latin scripts.

  8. multivariateGPT: a decoder-only transformer for multivariate categorical and numeric data

    cs.LG 2025-05 conditional novelty 6.0 of 10

    multivariateGPT extends next-token prediction to jointly predict the class and continuous value of mixed categorical and numeric time series, with Gaussian uncertainty, and outperforms discrete-token baselines on clin...

  9. BALM-TSF: Balanced Multimodal Alignment for LLM-Based Time Series Forecasting

    cs.AI 2025-08 conditional novelty 5.0 of 10

    BALM-TSF combines a statistical-prompt text branch with a patch-based time series branch, using scaling plus contrastive alignment to balance the two modalities, improving long-term and few-shot forecasting on five of...

  10. Improving Large Language Models with Concept-Aware Fine-Tuning

    cs.LG 2025-06 conditional novelty 5.0 of 10

    Adding lightweight multi-token auxiliary heads with a weighted future-token loss improves supervised fine-tuning of Llama-3-8B-Instruct across five diverse tasks.

  11. From Data to Device: ELMOD An Efficient German-First 2.7B Language Model for Mobile Inference

    cs.CL 2026-07 conditional novelty 4.0 of 10

    A 2.7B German-first LLM trained cheaply on public data with language-specific quality filtering matches larger 7B models on German reasoning benchmarks and runs on-device.

Pith tools