Multi-Pass Sentence Cloze Infilling with Sigmoidal Semantic Congruence for Robust AI Text Detection

Abstract

The rapid proliferation of Large Language Models (LLMs) has created an urgent imperative for reliable, domain-general machine-generated text detection. Existing statistical detectors relying on token-level perplexity or white-box log-likelihoods exhibit pronounced vulnerabilities, frequently misclassifying structured, formal human writing and suffering from elevated False Positive Rates (FPR). In this paper, we introduce <strong>ClozeCongruence</strong>, a black-box detection framework grounded in bidirectional contextual sentence predictability across multi-passage discourse. Rather than inspecting token logits, our method executes a dual-pass cloze protocol: (1) an alternate-sentence cloze pass evaluating structural continuity, and (2) a centroid three-sentence extraction pass measuring central argument reconstructability. Multi-sentence infilled hypotheses are dynamically aligned using Max-Weight Bipartite Hungarian Permutation Matching (S3) to eliminate arbitrary sentence-ordering penalties. To evaluate semantic alignment without step discontinuities, we introduce Continuous Sigmoidal Dynamic Gating (k=15, c0=0.70) that seamlessly interpolates between DeepEval propositional meaning overlap and semantic cosine embeddings. Across extensive empirical evaluations on 240 multi-passage essays across six academic disciplines (Cognitive Neuroscience, Quantitative Economics, Distributed Systems, Philosophy of Mind, Molecular Genetics, and Modern History), our framework achieves 99.58% accuracy, 99.17% precision, and a strictly bounded 0.83% False Positive Rate on authentic human writing.