An Interactive Agentic Framework for Deep Knowledge Extraction

Paper (arXiv) | Project Page | Code

Overview

This framework systematically extracts and quantifies the latent knowledge of black-box LLMs through interactive agentic exploration.

Accepted to the 43rd International Conference on Machine Learning (ICML 2026).

Installation

# Clone the repository
git clone https://github.com/ulab-uiuc/KnowledgeExtraction.git
cd KnowledgeExtraction

# Create conda environment
conda create -n ke python=3.10 -y
conda activate ke

# Install dependencies
pip install -r requirements.txt

Prepare API Key(s)

We use NVIDIA NIM API for LLM inference. Follow these steps to get your API key:

1. Register: Go to NVIDIA NIM and create an account
2. Get API Key: Navigate to the API Keys section and create a new key
3. Create config file: Create api.json in the project root:

{
    "api_keys": [
        "nvapi-xxxxxxx-your-first-key-xxxx",
        "nvapi-yyyyyyy-your-second-key-yyyy"
    ],
    "embed_config": {
        "model": "Qwen/Qwen3-Embedding-8B",
        "base_url": "http://localhost:30000/v1"
    }
}

Tip: You can add multiple API keys (from multiple accounts) to distribute the load and avoid rate limits.

Launch Embedding Server (Optional)

For better deduplication performance, launch a local embedding server:

bash launch_sglang.sh Qwen/Qwen3-Embedding-8B 30000 0.0.0.0 0

Quick Start

python main.py --domain "Deep Learning" --model "meta/llama-3.1-70b-instruct"

Reproduce Experiments

We provide scripts to reproduce all experiments from the paper:

Experiment	Description	Command
Exp1: Strategy Search	Pareto frontier analysis	python scripts/run_pareto_curves.py
Exp2: Scaling Law	Cross-scale comparison (8B/70B/405B)	python scripts/run_size_comparison.py
Exp3: Specialization	General vs RL-tuned models	python scripts/run_evolution_comparison.py
Exp4: Cross-Series	Different model families (~7B)	python scripts/run_cross_series.py

Example: Reproduce Experiment 1

# Run extraction with all strategies
python scripts/run_pareto_curves.py

# Evaluate and generate plots
python scripts/evaluate_pareto_curves.py
python scripts/plot_aggregated_pareto.py

Project Structure

KnowledgeExtraction/
├── agents/                 # LLM client and API handling
│   ├── call_agent.py      # Generation wrapper with token tracking
│   └── clientpool.py      # Multi-key rotation and retry logic
├── core/                   # Core processing modules
│   ├── processor.py       # Embedding and deduplication
│   ├── judge.py           # Domain relevance auditing
│   ├── evaluator.py       # Evaluation metrics computation
│   └── cleaner.py         # Bullet point extraction
├── pipelines/              # Exploration strategies
│   ├── base.py            # Base pipeline with saturation detection
│   ├── p2_sequential.py   # Sequential probing
│   ├── p3_reflection.py   # Self-reflective refinement
│   ├── p4_taxonomy_explorer.py  # Recursive taxonomy
│   └── p5_debate.py       # Multi-perspective debate
├── scripts/                # Experiment scripts
│   ├── run_*.py           # Experiment runners
│   ├── evaluate_*.py      # Evaluation pipelines
│   └── plot_*.py          # Visualization scripts
└── utils/                  # Utility functions

Citation

If you find this work useful, please cite our paper:

@inproceedings{yang2026probing,
  title={Probing the Knowledge Boundary: An Interactive Agentic Framework for Deep Knowledge Extraction},
  author={Yuheng Yang and Siqi Zhu and Tao Feng and Ge Liu and Jiaxuan You},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year={2026},
  eprint={2602.00959},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.00959}
}

License

This project is licensed under the MIT License.