Research map / Learning theory and algorithms / Combinatorial and search optimization
The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks
Ziqian Zhong, Ziming Liu, Max Tegmark, Jacob Andreas
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 |
|---|---|---|---|
| model with constant attention (no attention) uses Pizza instead of Clock formula to explain logits | Model A · modular addition p=59 | fraction of variance explained: 75.41 -> 99.18 | Improved |
| keep three circles (6 PCs) instead of one circle (2 PCs) in embeddings (ablation) | Model A · modular addition circle isolation | overall accuracy: 32.8 -> 91.4 | Improved |
| add accompanying pizzas (circles #4-#6) to the embedding (ablation) | Model A · modular addition | accuracy: 99.7 -> 100 | Improved |
| increase attention rate (attention strength) in one-layer transformer (ablation) | one-layer transformer width 128 · modular addition | algorithm used | Mixed |
| increase model width (ablation) | one-layer transformer · modular addition | attention-rate phase boundary | Mixed |
| increase network depth to 2-4 layers (ablation) | transformers · modular addition | fraction of circular models: 34.31% (1 layer) -> 9.95%, 11.55%, 6.08% | Hurt |
Closest papers by meaning
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.87)
- Learning to Add, Multiply, and Execute Algorithmic Instructions Exactly with Neural Networks (NeurIPS 2025 · similarity 0.87)
- Emergence in non-neural models: grokking modular arithmetic via average gradient outer product (ICML 2025 · similarity 0.87)
- Feature emergence via margin maximization: case studies in algebraic tasks (ICLR 2024 · similarity 0.86)
- Transformers Converge to Invariant Algorithmic Cores (NeurIPS 2026 · similarity 0.86)
- On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods (ICLR 2024 · similarity 0.86)
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.
