Reproduce “FoNE: Precise Single-Token Number Embeddings via Fourier Features” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianyi Zhou, Deqing Fu, Mahdi Soltanolkotabi, Robin Jia, Vatsal Sharan
Researchers at the University of Southern California introduce FONE (Fourier Number Embedding), a method to represent numbers as precise, single-token Fourier features within Large Language Models, achieving perfect or near-perfect accuracy on various arithmetic tasks while significantly reducing data and parameter requirements compared to conventional tokenization.
Key insights
- Leveraging Fourier features with a multi-period design enables FONE to capture exact numerical properties of numbers, preserving numeracy and full digit recovery robustly against normalization.
- Representing numbers as single, precise tokens eliminates frequency bias and fragmentation issues, leading to substantial gains in data and parameter efficiency for numerical tasks.
- FONE's integration improves LLM arithmetic performance without compromising general language modeling capabilities, even reducing validation perplexity on text tokens containing numbers during pre-training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.