Full State-Based CS-QTM: An Architecture for Question Intervention Control in LLM-Based Chatbots

Questions in large language model (LLM)-based chatbots can elicit information, but unnecessary questions create over-intervention, and omissions create under-intervention. We formulate question timing as a control problem and present Full State-Based CS-QTM. The architecture combines Episode Memory, accumulated question state qt, readinesst, question_scoret, adaptive_thresholdt, Prompt Routing, Hard Reset, Cooldown, and Topic Shift handling. A threshold was selected on a conversation-disjoint development set and frozen before evaluation. Evaluation covered 100 Korean dialogue scenarios and 645 turns, with 180 question-mark-aligned Ask and 465 Not Ask operational references. Qwen Frozen Full obtained F1 = 0.3226 versus 0.2332 for the same-generator Local LLM-only; a shallow sequential classifier trained on development data obtained F1 = 0.6139, revealing surface and sequential regularities in labels. Without retuning, EXAONE Frozen Full and LLM-only obtained F1 = 0.3024 and 0.2865. Controlled comparisons showed that threshold policy and generated-context forcing shifted the over-/under-intervention balance; grouped linguistic-feature effects remained uncertain after multiplicity correction. Internal logs separated policy, visible realization, resets, cooldown, topic shifts, and FP/FN paths. These results permit inspection of the closed-loop control process; they do not establish universal model superiority, optimal timing, or benchmark state-of-the-art.