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 boundaries extend. 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 granularities. To ensure the quality of extracted knowledge, we introduce a three-stage knowledge processing pipeline that combines vector-based filtering to remove exact duplicates, LLM-based adjudication to resolve ambiguous semantic overlaps, 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 extract more knowledge. In addition, we identify a Pass@1-versus-Pass@k trade-off: domain-specialized models achieve higher initial accuracy but degrade rapidly, while general-purpose models maintain stable performance during extended extraction. Finally, our results show that differences in training data composition lead to distinct and measurable knowledge profiles across model families.

💡 Research Summary

The paper tackles the problem of quantifying what large language models (LLMs) actually know—a “knowledge boundary”—by moving beyond static benchmark evaluations toward an interactive, saturation‑based probing methodology. The authors introduce an Interactive Agentic Framework that consists of two main components: (1) a suite of four adaptive exploration policies that actively interrogate a black‑box LLM, and (2) a three‑stage Knowledge Processor that rigorously removes duplicates, adjudicates semantic overlap, and audits domain relevance.

Exploration policies

1. Sequential Associative Probing – a baseline that repeatedly asks “What else?” or “Give more details,” relying only on the model’s associative memory.
2. Self‑Reflective Refinement – a critic‑actor loop where the model first audits its own prior outputs, identifies gaps or inconsistencies, and then generates targeted prompts to fill them.
3. Recursive Taxonomy Explorer – the most sophisticated strategy. It decomposes a target topic into a hierarchical tree using Bloom’s taxonomy (fact, concept, procedural) and then spawns parallel agents at leaf nodes. By controlling branching factor (W) and maximum depth (Dmax), the method forces the model to surface long‑tail facts that would otherwise be suppressed by dominant high‑probability tokens.
4. Multi‑Perspective Parallel Probing – creates N distinct expert personas (engineer, legal scholar, ethicist, etc.) and lets each independently contribute insights, thereby capturing niche knowledge that a single neutral agent might miss.

Knowledge Processor
Stage 1: Vector Filtering – embeddings from qwen3‑8b‑emb are used; pairs with cosine similarity > 0.92 are merged instantly.
Stage 2: LLM‑Based Adjudication – for ambiguous pairs (0.70 < sim < 0.92), a stronger LLM (DeepSeek‑V3.1) acts as a judge to decide whether the statements describe the same core fact, handling negations and subtle technical differences that pure vector similarity cannot catch.
Stage 3: Domain Relevance Auditing – each remaining atom is evaluated against Bloom’s taxonomy criteria to keep only factual, conceptual, or procedural knowledge, discarding meta‑statements, generic fluff, or incomplete fragments.

Metrics and Saturation Criteria
The framework tracks per‑turn novel atoms (nₜ), raw atoms (rₜ), growth rate gₜ = nₜ/|K₁:ₜ₋₁|, and efficiency eₜ = nₜ/rₜ. Extraction stops when any of the following hold: gₜ < 1 %, eₜ < 10 %, nₜ < 3, or 15 turns are reached. Yield (Yₜ) is measured as the ratio of unique valid atoms to a baseline (the final set from the Recursive Taxonomy (L2W2) strategy). Cost‑yield curves (Cₜ, Yₜ) visualize Pareto efficiency. For cross‑model comparison, Recall = |K_model| / |K_union| (where K_union aggregates all models’ valid atoms) and Accuracy = |K_valid| / |K_unique| are reported.

Experimental Findings

• Recursive Taxonomy consistently outperforms the other three policies, achieving the highest yield at the lowest token cost, thus occupying the top‑left of the Pareto frontier.
• A Knowledge Scaling Law emerges: larger models extract more unique atoms, with growth that is roughly linear to super‑linear relative to parameter count.
• Pass@1 vs Pass@k trade‑off: domain‑specialized fine‑tuned models show high initial accuracy (Pass@1) but their recall deteriorates sharply as extraction proceeds, whereas general‑purpose models maintain steadier recall across many turns.
• Training data composition shapes distinct knowledge profiles; models from different families (e.g., OpenAI vs. Meta) exhibit divergent recall patterns on the same topic, confirming that data distribution influences the shape of the knowledge boundary.

Implications
By treating knowledge extraction as a saturation‑driven tree search, the framework reveals latent knowledge that static tests miss, providing a quantitative map of an LLM’s usable knowledge space. This has immediate relevance for model interpretability, alignment, safety auditing, and for building downstream knowledge bases that rely on verified factual content. The authors suggest future directions such as meta‑learning of exploration policies, multimodal domain auditing, and automatic structuring of extracted atoms into ontologies.

In summary, the paper delivers a novel, systematic methodology for probing and measuring the knowledge limits of black‑box LLMs, demonstrating that adaptive agentic exploration combined with rigorous semantic deduplication yields a scalable, reproducible, and insightful view of what modern language models truly know.