REVIEW 14 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
Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs
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.
Forward citations
Cited by 14 Pith papers
-
The Shape of Addition: Geometric Structures of Arithmetic in Large Language Models
LLM residual streams during addition form an Iso-Raw-Sum Trajectory anchored by digit semantics and modulated by continuous carry signals, with errors arising as geometric slippages across quantization thresholds in a...
-
Large Language Models as Amortized Pareto-Front Generators for Constrained Bi-Objective Convex Optimization
DIPS fine-tunes LLMs to output ordered feasible decision vectors approximating Pareto fronts for constrained bi-objective convex problems, reaching 95-98% normalized hypervolume with 0.16s inference.
-
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
Subword tokenization impairs phonological knowledge encoding in LMs, but an IPA-based fine-tuning method restores it with minimal impact on other capabilities.
-
Efficient numeracy in language models through single-token number embeddings
BitTokens represent numbers as single tokens via IEEE 754 binary format, allowing small language models to learn basic arithmetic algorithms nearly perfectly.
-
FLEXITOKENS: Flexible Tokenization for Evolving Language Models
FLEXITOKENS replaces rigid subword tokenizers and fixed-compression auxiliary losses with a simplified boundary-prediction objective in byte-level models, yielding lower over-fragmentation and up to 10-point gains on ...
-
Variable-length Gray codes for the Natural Numbers
Every finite digit string gets exactly one number, and consecutive numbers differ by a single edit — a complete, locally stable variable-length Gray code.
-
Inside the LLM Word Factory
Activation patching localizes English detokenization in Llama2-7B to a two-stage attention-then-MLP process at layer 1 that generalizes to 12 models from 8 families, with depth varying by positional encoding, plus an ...
-
The Effect of Scripts and Formats on LLM Numeracy
LLM arithmetic accuracy falls sharply when numerals leave the familiar Hindu–Arabic format, and few-shot prompting with examples narrows most of that gap.
-
MachineLearningLM: Scaling Many-shot In-context Learning via Continued Pretraining
MachineLearningLM uses continued pretraining on SCM-synthesized ML tasks with random-forest distillation to give LLMs robust many-shot in-context learning on tabular classification, reaching random-forest accuracy lev...
-
The Importance of Being Statistically Earnest: A Critical Re-evaluation of GSM-Symbolic
Re-evaluation of GSM-Symbolic using GLMMs on 20 models shows only half have significant performance changes; a distribution shift in larger integers (K-S=0.12) accounts for significance in half the remaining cases.
-
Understanding Secret Leakage Risks in Code LLMs: A Tokenization Perspective
BPE tokenization creates gibberish bias in CLLMs, causing secrets with high character entropy but low token entropy to be preferentially memorized due to training data distribution shifts.
-
A Triadic Suffix Tokenization Scheme for Numerical Reasoning
Triadic Suffix Tokenization groups digits into triads with fixed magnitude suffixes to make order-of-magnitude relationships explicit at the token level for LLMs.
-
BALM-TSF: Balanced Multimodal Alignment for LLM-Based Time Series Forecasting
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...
-
From Data to Device: ELMOD An Efficient German-First 2.7B Language Model for Mobile Inference
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.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.