Not All Weights Are Equal
Abstract
We test whether a dedicated ternary-weight reasoning component outperforms a homogeneous MLP baseline on multi-step logic problems requiring generalization to unseen entities. In two runs, the reasoning component outperformed the baseline by 12.2% and 19.6% respectively, while the baseline overfit and stalled. We propose a full architecture built on this result, with separate components for reasoning and language.
Fields of Study & Subjects
Topic Modeling
Logic
Reasoning
and Knowledge
Multimodal Machine Learning Applications
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 5 March 2026
- Language
- EN