Same logic, different words: our NeurIPS 2026 paper on why LLMs contradict themselves

Our paper “When Form Changes but Logic Doesn’t: Building Logic-invariant LLMs through Structures” (with Xuyuan Liu, Xinshuai Dong, and Yujun Yan) has been accepted at NeurIPS 2026.

Ask a large language model the same rule-application question phrased two logically equivalent ways, and strong models flip their answers on over a third of problems — a reliability gap that benchmark accuracy alone never reveals. We trace the failure to models reading wording rather than logic, show that prompting-based fixes barely help, and propose a structural remedy: encoding each problem’s logic as a decision diagram and training the model to consult that wording-independent representation while it reasons. The approach roughly halves self-contradiction, improves accuracy most on logical forms never seen in training, and transfers to expert-written reasoning benchmarks — at negligible added cost.

🔗 Links:

📝 Blog post for a broader audience
📄 arXiv:2610.25402
🏛️ NeurIPS 2026 page
💻 Code on GitHub