Symmetry-Aware Foundation Model for Logic Rule Induction
A recent study published on arXiv (2608.00383) presents a foundation model for logical rule induction that is aware of symmetries, including atom naming, example order, polarity flips, and label swaps. This model, based on the Neural Rule Inducer (a disjunctive-normal-form foundation model), achieves precise equivariance through a canonical export that translates discrete rules from literal scores without the need for retraining. In synthetic stress tests, the accuracy for support labels remains consistent with larger schemas, while rule fidelity on new inputs surpasses that of the original model. Notably, accuracy improves significantly on larger schemas when applied to real data. The paper serves as a cross-type announcement, suggesting it may have been shared in other venues. This research is pertinent to AI and machine learning, especially in interpretable rule learning and symmetry-aware frameworks.
Key facts
- Paper arXiv:2608.00383 introduces a symmetry-aware foundation model for logic rule induction.
- The model respects symmetries: atom naming, example order, polarity flips, and label swap.
- A canonical export decodes discrete rules from literal scores, requiring no retraining.
- The model is exactly equivariant when scores respect the symmetries.
- Instantiated on the Neural Rule Inducer, a DNF foundation model that natively respects only example order.
- Remaining symmetries are restored through architecture, inference, and training.
- On synthetic stress tests, accuracy on support labels stays stable at larger schemas.
- Rule fidelity on fresh inputs remains above the unmodified model.
- On real data, accuracy improves most on larger schemas.
- The paper is a cross-type announcement on arXiv.
Entities
Institutions
- arXiv