Research map / Large language models / Language, safety and interpretability
FoNE: Precise Single-Token Number Embeddings via Fourier Features
Tianyi Zhou, Deqing Fu, Mahdi Soltanolkotabi, Robin Jia, Vatsal Sharan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use single-token FoNE instead of digit-wise tokens | transformer · 3-digit multiplication | accuracy (%): 81.21 -> 98.56 | Improved |
| use single-token FoNE instead of subword tokens | transformer · 3-digit multiplication | accuracy (%): 8.05 -> 98.56 | Improved |
| use FoNE instead of xVal | transformer · 6-digit decimal addition | accuracy (%): 0.44 -> 100 | Improved |
| use FoNE instead of digit-wise tokens | transformer · 6-digit decimal addition | training time: 11'48'' -> 3'18'' | Improved |
| use FoNE (one token) instead of digit-wise | transformer · 6-digit decimal addition | tokens per number: 7 -> 1 | Improved |
| replace linear layer with zero padding to match embedding dimension (ablation) | FoNE · Multiplication | accuracy (%): 99.95 -> 99.91 | No clear effect |
| add periods 2 and 5 to period 10 (ablation) | FoNE · Multiplication | accuracy (%): 99.95 -> 99.99 | No clear effect |
| use period 5 instead of 10 (ablation) | FoNE · Decimal addition | accuracy (%): 100 -> 1.52 | Hurt |
| use period 7 instead of 10 (ablation) | FoNE · Decimal addition | accuracy (%): 100 -> 3.64 | Hurt |
| add simplified FoNE to Llama-3.1-1B continual pretraining | Llama-3.1-1B · 4-digit subtraction | accuracy (%): 29.60 -> 39.90 | Improved |
| add simplified FoNE to Llama-3.1-1B | Llama-3.1-1B · MMLU | accuracy (%): 38.10 -> 38.21 | No clear effect |
| use FoNE tokenization in GPT-2 117M pretraining | GPT-2 117M · FineWeb text-token perplexity | perplexity: 67.390 -> 46.856 | Improved |
Later papers that compared with this method
1 later accepted paper reports FoNE as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- GeoNum: Bridging Numerical Continuity and Language Semantics via Geometric Embedding (AAAI 2026 · better on 8 of 8 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Efficient numeracy in language models through single-token number embeddings (ICML 2026 · similarity 0.88)
- Regress, Don't Guess: A Regression-like Loss on Number Tokens for Language Models (ICML 2025 · similarity 0.87)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.85)
- Rep2Text: Decoding Full Text from a Single LLM Token Representation (NeurIPS 2026 · similarity 0.85)
- Scaling Embedding Layers in Language Models (NeurIPS 2025 · similarity 0.85)
- MathPile: A Billion-Token-Scale Pretraining Corpus for Math (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
