Not All Weights Are Equal

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.