Large Language Models cannot reliably distinguish what they know from what theydon’t—leading to hallucination rather than acknowledgment of uncertainty. Furthermore,once trained, model knowledge is frozen; incorporating new information requires expensiveretraining. We propose Bayesian-Gated Selective Learning (BGSL), a frameworkenabling controlled post-deployment learning. BGSL employs two complementary trigger mechanisms: (1) token-level uncertainty signalsthat fire immediately when the model encounters unfamiliar concepts, and (2) usagepatterndetection that accumulates evidence when familiar terms are used in ways that don’tmatch the model’s understanding. When either trigger fires, the system simultaneously retrievesexternal information to ground the immediate response and packages the interactionfor deferred training. Learning occurs in audited batch cycles, enabling human review andsafety verification. Over successive deployment cycles, the model genuinely internalizes newknowledge, reducing reliance on repeated retrieval for stable information. We ground this framework in Bayesian-inspired reasoning—treating uncertainty as aproxy for prior strength—as a design principle rather than a formal implementation. Thispaper details the conceptual architecture, anticipated challenges, and evaluation protocols.Future work must determine the specific signal implementations and thresholds appropriatefor production systems.Note: This work is a conceptual framework and does not present empirical results.
Edward Dean Brown (Wed,) studied this question.