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

• URL: http://arxiv.org/abs/2402.14903v1
• Date: Thu, 22 Feb 2024 18:14:09 GMT
• Title: Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs
• Authors: Aaditya K. Singh, DJ Strouse
• Abstract summary: Tokenization is the division of input text into input tokens. We study the effect this choice has on numerical reasoning through the use of arithmetic tasks.
• Score: 3.6722413665749674
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Regress, Don't Guess -- A Regression-like Loss on Number Tokens for Language Models [2.5346260093097017]
  We present two versions of a number token loss for language models. The first is based on an $L_p$ loss between the ground truth token value and the weighted sum of the predicted class probabilities. The second loss minimizes the Wasserstein-1 distance between the distribution of the predicted output probabilities and the ground truth distribution.
  arXiv  Detail & Related papers  (2024-11-04T13:43:24Z)
• Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles [23.134664392314264]
  Tokenization is associated with many poorly understood shortcomings in language models (LM) This work studies how tokenization impacts model performance by analyzing and comparing models with their byte-level counterparts. We develop a next-byte sampling algorithm that eliminates tokenization bias without requiring further training or optimization.
  arXiv  Detail & Related papers  (2024-10-11T23:30:42Z)
• Subtle Errors Matter: Preference Learning via Error-injected Self-editing [59.405145971637204]
  We propose a novel preference learning framework called eRror-Injected Self-Editing (RISE) RISE injects predefined subtle errors into partial tokens of correct solutions to construct hard pairs for error mitigation. Experiments validate the effectiveness of RISE, with preference learning on Qwen2-7B-Instruct yielding notable improvements of 3.0% on GSM8K and 7.9% on MATH.
  arXiv  Detail & Related papers  (2024-10-09T07:43:38Z)
• Graph-Structured Speculative Decoding [52.94367724136063]
  Speculative decoding has emerged as a promising technique to accelerate the inference of Large Language Models. We introduce an innovative approach utilizing a directed acyclic graph (DAG) to manage the drafted hypotheses. We observe a remarkable speedup of 1.73$times$ to 1.96$times$, significantly surpassing standard speculative decoding.
  arXiv  Detail & Related papers  (2024-07-23T06:21:24Z)
• Understanding and Mitigating Tokenization Bias in Language Models [6.418593476658017]
  State-of-the-art language models are autoregressive and operate on subword units known as tokens. We show that popular encoding schemes induce a sampling bias that cannot be mitigated with more training or data. We propose a novel algorithm to obtain unbiased estimates from any language model trained on tokenized data.
  arXiv  Detail & Related papers  (2024-06-24T17:38:02Z)
• Improving Input-label Mapping with Demonstration Replay for In-context Learning [67.57288926736923]
  In-context learning (ICL) is an emerging capability of large autoregressive language models. We propose a novel ICL method called Sliding Causal Attention (RdSca) We show that our method significantly improves the input-label mapping in ICL demonstrations.
  arXiv  Detail & Related papers  (2023-10-30T14:29:41Z)
• Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks [54.306234256074255]
  We identify the issue of tokenization inconsistency that is commonly neglected in training generative models. This issue damages the extractive nature of these tasks after the input and output are tokenized inconsistently. We show that, with consistent tokenization, the model performs better in both in-domain and out-of-domain datasets.
  arXiv  Detail & Related papers  (2022-12-19T23:33:21Z)
• Induced Natural Language Rationales and Interleaved Markup Tokens Enable Extrapolation in Large Language Models [8.166629393064097]
  The ability to extrapolate, i.e., to make predictions on sequences that are longer than those presented as training examples, is a challenging problem for deep learning models. Recent work shows that this limitation persists in state-of-the-art Transformer-based models. We demonstrate that large language models can succeed in extrapolation without modifying their architecture or training procedure.
  arXiv  Detail & Related papers  (2022-08-24T11:25:27Z)
• When Counting Meets HMER: Counting-Aware Network for Handwritten Mathematical Expression Recognition [57.51793420986745]
  We propose an unconventional network for handwritten mathematical expression recognition (HMER) named Counting-Aware Network (CAN) We design a weakly-supervised counting module that can predict the number of each symbol class without the symbol-level position annotations. Experiments on the benchmark datasets for HMER validate that both joint optimization and counting results are beneficial for correcting the prediction errors of encoder-decoder models.
  arXiv  Detail & Related papers  (2022-07-23T08:39:32Z)
• You should evaluate your language model on marginal likelihood overtokenisations [5.824498637088864]
  We argue that language models should be evaluated on their marginal likelihood over tokenisations. We evaluate pretrained English and German language models on both the one-best-tokenisation and marginal perplexities.
  arXiv  Detail & Related papers  (2021-09-06T15:37:02Z)
• ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators [108.3381301768299]
  Masked language modeling (MLM) pre-training methods such as BERT corrupt the input by replacing some tokens with [MASK] and then train a model to reconstruct the original tokens. We propose a more sample-efficient pre-training task called replaced token detection.
  arXiv  Detail & Related papers  (2020-03-23T21:17:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.