Reproduce “Grokking Group Multiplication with Cosets” 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
Dashiell Stander, Qinan Yu, Honglu Fan, Stella Biderman
EleutherAI researchers found that a neural network grokking group multiplication on symmetric groups S5 and S6 implements the operation through "coset circuits," which precisely leverage the mathematical structure of group cosets. Rigorous causal interventions confirmed this mechanism and showed prior interpretations for the same problem were incomplete or misleading, advocating for higher standards in mechanistic interpretability.
Key insights
- Neural networks can discover and leverage exact mathematical structures, specifically group cosets, to solve algebraic problems like group multiplication.
- The phenomenon of grokking, where a model transitions from memorization to generalization, involves the formation of clean, generalizable circuits aligned with underlying mathematical algorithms.
- Rigorous causal intervention is essential for validating mechanistic interpretations, as observational evidence can lead to incomplete or misleading conclusions, requiring a higher standard for interpretability research.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.