Phantom Paradoxes: Language Models Treat Unresolved Circular Truth Dependence as Contradiction
The study found that large language models are perturbed not by general self-reference, but by 'non-closing truth recursion' (NCTR)—situations where truth values cannot be settled in a finite number of steps. This instability is linked to undecidable problems in matrix theory, suggesting models treat these unresolved circular dependencies as contradictions.