CoTu at EXACT 2026: Neuro-Symbolic Reasoning for Transparent Educational QA

Small models can provide transparent reasoning: The winning system used only a 4B parameter model (much smaller than large proprietary systems like GPT-4), demonstrating that explainability doesn't require massive models—a key finding for making AI more accessible and auditable.

Symbolic verification improves trustworthiness: By grounding answers in formal logic solvers (Z3) for regulatory questions and numerical Python for physics problems, the system made its reasoning verifiable and checkable by humans, rather than relying on opaque neural network outputs alone.

Self-correction loops enhance reliability: The system incorporated a shared self-correction mechanism where the model could review and improve its own reasoning, achieving a perfect score on physics problems—suggesting that built-in verification processes can significantly boost accuracy in high-stakes educational contexts.

Explainability and performance aren't trade-offs: The system achieved the highest technical score (13.44/15) while providing natural-language explanations for every answer, contradicting the assumption that transparency requires sacrificing quality—important for responsible AI deployment in education.

Practical constraints matter for real-world AI: The system had to meet a 60-second response limit while self-hosting on modest hardware, showing that responsible AI systems must balance transparency, accuracy, and computational efficiency to be genuinely deployable in practice.