Symmetry and Optimization in Neurosymbolic Reasoning Shortcuts
A recent paper on arXiv (2608.10420) investigates reasoning shortcuts in neurosymbolic systems, where unintended concepts lead to accurate predictions. The authors critique a framework proposed by Takemura, Inoue, and Nishino, which employs automorphism groups of value relabelings to determine when rules clarify concepts. Their findings indicate that the framework's primary definition—a uniform permutation applied across all positions—fails to hold for any of the four diverse benchmarks tested. The straightforward method of padding domains to a uniform size results in misleading pathology: 90.91% of solution pairs on CLE4EVR were found unexplained, while every clearly defined member of their new hierarchy shows 0%. Additionally, the content of the padded verdict varies with configuration-file ordering. Upon reassessing eleven rule families against fifteen predetermined predictions (thirteen confirmed), unexplained-pair rates range from 0% to 99%. The paper proposes a new hierarchy to tackle these challenges and offers a refined analysis of reasoning shortcuts.
Key facts
- Paper arXiv:2608.10420 examines reasoning shortcuts in neurosymbolic systems.
- Critiques Takemura, Inoue, and Nishino's framework using automorphism groups.
- Framework's key definition does not apply to any of the four benchmarks.
- Padding domains to a common size produces 90.91% unexplained pairs on CLE4EVR.
- Introduced hierarchy reports 0% unexplained pairs for well-defined members.
- Padded verdict's content rotates with configuration-file ordering.
- Re-measuring eleven rule families under fifteen predictions, thirteen confirmed.
- Unexplained-pair rates range from 0% to 99%.
Entities
Institutions
- arXiv