Reproduce “CALT: A Library for Computer Algebra with Transformer” 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
Hiroshi Kera, Shun Arakawa, Yuta Sato
CALT is a Python library designed to simplify the training of Transformer models for symbolic computation tasks, addressing the technical barrier faced by computer algebraists. It enables models to achieve a 74.5% success rate for integer factorization and an 81.2% success rate for polynomial multiplication over Z[x,y], with Chain-of-Thought reasoning improving modular arithmetic performance from 31.7% to 74.2% over F7[x,y].
Key insights
- Chain-of-Thought (CoT) reasoning substantially improves Transformer performance in challenging symbolic domains, such as modular arithmetic, by explicitly guiding models through intermediate steps.
- Transformers demonstrate strong capabilities in learning integer factorization and polynomial multiplication over integers, but struggle with modular arithmetic without explicit reasoning guidance.
- The library facilitates empirical studies on Transformer learning dynamics in symbolic contexts, revealing distinct convergence patterns across different task types (e.g., steady vs. stagnated convergence).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.