LLMs Know More About Numbers than They Can Say
Fengting Yuchi Li Du Jason Eisner · 2026
Read LLMs Know More About Numbers than They Can Say by Fengting Yuchi Li Du Jason Eisner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07812.txt.
Description
Language models stumble when comparing numbers written in different notations, like scientific notation against plain decimals. Fengting Yuchi and two co-authors ask whether the models secretly know the magnitudes even when their spoken answers go wrong.
Probing the hidden states of smaller open models, they find a single linear projection recovers log-magnitudes with about 2.3 percent relative error on synthetic text and 19 percent on scientific papers, and a linear classifier ranks paired numerals with over 90 percent accuracy. Yet the models' explicit answers score only 50 to 70 percent. Adding the probe's loss during fine-tuning lifts verbal accuracy by 3.22 percent, showing internal knowledge can be coaxed into speech.
Catalog record
| CREATOR | Fengting Yuchi Li Du Jason Eisner |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-07812