metadata
task_categories:
- text-classification
Numeracy Probing (arXiv Data)
This repository contains the arXiv data used in the paper LLMs Know More About Numbers than They Can Say.
The dataset is derived from allenai/peS2o.
Links
- Paper: LLMs Know More About Numbers than They Can Say
- GitHub Repository: VCY019/Numeracy-Probing
Dataset Summary
This dataset contains a subset of approximately 100k samples from scientific papers (arXiv) used to evaluate how Large Language Models (LLMs) represent numerical magnitudes. The associated study probes the hidden states of LLMs to determine if they encode the log-magnitudes and rankings of numerals, even when the models fail to compare them correctly in verbalized text.
The data was used specifically for training and evaluating linear probes (regression and classification) on scientific text as described in the EACL 2026 paper.
Citation
@inproceedings{yuchi-du-eisner-2026,
author = {Fengting Yuchi and Li Du and Jason Eisner},
title = {{LLM}s Know More About Numbers than They Can Say},
booktitle = {Proceedings of the Conference of the European Chapter
of the Association for Computational Linguistics: Human
Language Technologies (EACL)},
year = {2026},
month = mar,
address = {Rabat, Morocco},
note = {Oral presentation.},
url = {https://arxiv.org/abs/2602.07812}
}