Research map / Label-efficient and robust learning / Recognition and long-tail
Intrinsic Task Symmetry Drives Generalization in Algorithmic Tasks
Hyeonbin Hwang, Yeachan Park
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 |
|---|---|---|---|
| add symmetry-prompting auxiliary loss (commutativity/associativity, triangular symmetry, transitivity) (ablation) | Transformer on modular arithmetic, graph metric completion, comparison tasks · x+y mod p, lattice graph, 2D comparison | epochs to reach accuracy 1 | Improved |
| add geometry-prompting priors (nuclear norm, entropy, Lipschitz regularization) (ablation) | Transformer on algorithmic tasks · x+y mod p, lattice graph, 2D comparison | epochs to reach accuracy 1 | Improved |
| add GrokFast optimizer-based acceleration (ablation) | Transformer on algorithmic tasks · modular arithmetic, graph metric completion, comparison | epochs to reach accuracy 1 | Mixed |
| add symmetry-prompting loss to two-layer MLP without weight decay (ablation) | Two-layer MLP on modular addition (80% train fraction) · modular addition, no weight decay | time to generalization: up to two orders of magnitude reduction | Improved |
| remove weight decay (ablation) | Modular addition model · modular addition | test accuracy | No clear effect |
| penalize symmetry acquisition (ablation) | Algorithmic task models · algorithmic tasks (Appendix F) | generalization | Hurt |
| lower symmetry violation metric below soft threshold | Modular addition models · modular addition, 20 seeds | test accuracy: test accuracy reaches 1.0 | Improved |
Closest papers by meaning
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (NeurIPS 2026 · similarity 0.90)
- Parameter symmetries determine representational geometry in overparameterized nonlinear networks (NeurIPS 2026 · similarity 0.88)
- Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking (NeurIPS 2025 · similarity 0.88)
- A mathematical theory of balancing relational generalization and memorization (NeurIPS 2026 · similarity 0.87)
- Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.87)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · 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.
