AI · 8 min read · April 17, 2026
Modular Neural Networks Learn Three-Valued Logic Without Symbolic Solvers
THEIA demonstrates that dedicated domain engines enable neural networks to master Kleene three-valued logic and generalize compositionally to sequences 100x longer than training.
THEIA, a modular neural system, learns complete three-valued logic end-to-end and generalizes to sequences far longer than training data.
- — Four specialized engines (arithmetic, order, set, propositional) feed into a final logic module for Kleene K3 coverage.
- — Trained on 2M samples; achieves all 12 K3 rules in ~8 minutes across five independent runs.
- — Generalizes from 5-step sequences to 500-step evaluation at 99.97% accuracy via modular structure.
- — Flat MLPs collapse to chance-level performance by 50 steps; modularity is causally necessary.
- — Mechanistic probing shows upstream engines delay truth-value commitment until the logic boundary.
- — Transformer baseline reaches similar accuracy through different representational dynamics (contraction-then-expansion).
- — Activation patching confirms 100% causal control over OR and AND outputs via logic engine.
Frequently asked
- Kleene three-valued logic (K3) extends classical true/false reasoning to include a third state: unknown or undefined. This models real-world scenarios where information is incomplete or contradictory. THEIA learns K3 to demonstrate that neural networks can handle uncertainty without external symbolic solvers, using four domain-specific engines that converge on a final logic module.