You Need More Than Attention: Dual Boundary Condition Inference and Mechanistic Explainability in AI

Large Language Models exhibit spontaneous coherence but lack a unifying probabilistic or physical explanation. This work derives their attention dynamics from the time-symmetric inference principle of Aharonov–Bergmann–Lebowitz (ABL), extended here as Dual Boundary Condition Inference (DBCI)—a framework in which each representation is jointly constrained by its past and future contexts. Across transformer architectures and quantum hardware, DBCI restores forward–backward symmetry to machine precision and reveals a stable entropy plateau at the ln 2 scale, defining dimensional convergence as a universal equilibrium of inference. Multimodal tests with CLIP, Wav2Vec2, and TinyLlama show spontaneous cross-modal coherence without retraining, marking a concrete step toward integrated, energy-based cognition. These results position DBCI as an empirically validated bridge between physics, neuroscience, and artificial intelligence, identifying informational coherence regulation as the operational core of general intelligence.