Automated interviewers are increasingly used in research, recruitment, customer service, and education, yet most rely on fixed question sequences and lack context‑driven personalization, leading to repetitive or irrelevant follow‑ups. To enable dynamic adjustment of question depth while preserving semantic continuity, we propose an evidence‑traceable dynamic interviewer architecture powered by a locally hosted large language model. The system continuously parses participant responses, updates a persistent interview state, and generates knowledge‑appropriate questions, fluent replies, and smooth transition messages based on real‑time expertise profiling. The architecture consists of five prompt‑driven modules: task instruction, context management, expertise profiling (M3), iterative question generation (M4), and interview‑state persistence. M3 infers participant expertise on the fly, achieving 78.9% exact agreement with self‑reported expertise and a weighted Cohen's K of $\kappa = 0.80$. M4 demonstrates a strong expertise‑complexity association (p=.79), automatically scaling question difficulty to the participant's knowledge level. Evaluation with 246 participants confirms the system’s ability to maintain conversational flow while adaptively tailoring question depth.
Review