Probing the Knowledge Boundary: An Interactive Agentic Framework for Deep Knowledge Extraction

Large Language Models (LLMs) can be seen as compressed knowledge bases, but it remains unclear what knowledge they truly contain and how far their knowledge boundary extends. Existing benchmarks are mostly static and provide limited support for systematic knowledge probing. In this paper, we propose an interactive agentic framework to systematically extract and quantify the knowledge of LLMs. Our method includes four adaptive exploration policies to probe knowledge at different granularity. To ensure the quality of extracted knowledge, we introduce a three-stage knowledge processing pipeline that combines vector-based filtering to remove strict duplicates, LLM-based adjudication to resolve ambiguous semantic overlap, and domain relevance auditing to retain valid knowledge units. Through extensive experiments, we find that Recursive Taxonomy is the most effective exploration strategy. We also observe a clear knowledge scaling law, where larger models consistently recover more knowledge. In addition, we identify a Pass@1 versus Pass@k trade-off: domain-specialized models achieve higher initial accuracy but experience rapid degradation, while general-purpose models maintain stable performance over extended extraction. Finally, our results show that differences in training data composition lead to distinct and measurable knowledge profiles across model families, reflecting how pretraining shapes each model's parametric knowledge.

Large language models (LLMs) like ChatGPT have absorbed vast amounts of human knowledge during training, but we have surprisingly little ability to measure exactly what they know — or don't know. Standard tests only probe a small, fixed set of questions, missing the full depth of what a model has learned. This paper introduces an automated system that acts like an interviewer, systematically exploring what an AI model knows about a given topic by asking progressively deeper and more specific questions. Rather than stopping at surface-level answers, the system decomposes topics into finer and finer subtopics, forcing the model to reveal knowledge it would not volunteer unprompted. A dedicated processing pipeline then removes duplicate and low-quality responses, keeping only genuine, verifiable knowledge statements. Using this approach, we find that larger models consistently know more, that models trained for specialized tasks (such as coding) gain precision in their area but lose breadth elsewhere, and that different AI organizations' models carry distinct knowledge "fingerprints" reflecting what data they were trained on. This work provides a practical tool for auditing AI systems before deployment, helping developers and users understand what a model truly knows.