Research map / Architectures and efficiency / RNNs and normalization
The Geometry and Topology of Circuits: the Manifolds of Modular Addition
Gabriela Moisescu-Pareja, Gavin McCracken, Harley Wiltzer, Colin Daniels, Vincent Létourneau, Doina Precup, Jonathan Love
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 |
|---|---|---|---|
| switch from uniform attention (Pizza) to trainable attention (Clock) | modular addition transformer · layer-1 PCA geometry | variance explained by top 2 PCs | No clear effect |
| switch from uniform attention to trainable attention | modular addition transformer · phase alignment distribution (max activation) | MMD: MMD 0.0338 | No clear effect |
| concatenate embeddings (MLP-Concat) instead of adding (MLP-Add) | MLP on modular addition · layer-1 manifold | CKA to torus vs disc: disc CKA 0.707 vs torus CKA 0.994 | Improved |
| change architecture (MLP-Add, Attention 0.0/1.0, MLP-Concat) | modular addition networks · logit manifold | CKA to circle: 0.926 to 0.986 | No clear effect |
| increase depth from 1 to 2-3 layers | MLP-Add and Attention · Betti numbers per layer | topology | Mixed |
| use Zhong et al. diagonal-dependence as Pizza signature | Clock vs Pizza distinction · activation dependence on a-b | activation pattern | Hurt |
Closest papers by meaning
- Uncovering a Universal Abstract Algorithm for Modular Addition in Neural Networks (NeurIPS 2025 · similarity 0.86)
- DeepLayout: Learning Neural Representations of Circuit Placement Layout (ICML 2025 · similarity 0.86)
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.86)
- Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery (NeurIPS 2026 · similarity 0.85)
- Alignment of CNN and Human Judgments of Geometric and Topological Concepts (AAAI 2025 · similarity 0.85)
- Feature Learning beyond the Lazy-Rich Dichotomy: Insights from Representational Geometry (ICML 2025 · 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.
