Bayesian-Gated Selective Learning: A Framework for Post-Training Knowledge Acquisition in Large Language Models

  Large Language Models cannot reliably distinguish what they know from what they
don’t—leading to hallucination rather than acknowledgment of uncertainty. Furthermore,
once trained, model knowledge is frozen; incorporating new information requires expensive
retraining. We propose Bayesian-Gated Selective Learning (BGSL), a framework
enabling controlled post-deployment learning.
  BGSL employs two complementary trigger mechanisms: (1) token-level uncertainty signals
that fire immediately when the model encounters unfamiliar concepts, and (2) usagepattern
detection that accumulates evidence when familiar terms are used in ways that don’t
match the model’s understanding. When either trigger fires, the system simultaneously retrieves
external information to ground the immediate response and packages the interaction
for deferred training. Learning occurs in audited batch cycles, enabling human review and
safety verification. Over successive deployment cycles, the model genuinely internalizes new
knowledge, reducing reliance on repeated retrieval for stable information.
  We ground this framework in Bayesian-inspired reasoning—treating uncertainty as a
proxy for prior strength—as a design principle rather than a formal implementation. This
paper details the conceptual architecture, anticipated challenges, and evaluation protocols.
Future work must determine the specific signal implementations and thresholds appropriate
for production systems.

Note: This work is a conceptual framework and does not present empirical results.