AIGT Agent

This is the repository of paper Beyond Text-Intrinsic Features: An Agentic Framework for Evidence-Based AI-Generated Text Detection, including source code, the 4K benchmark, and results for an LLM-based AI-generated text detector that combines machine-learning detectors with external evidence from online search.

The main system uses MPU and Binoculars.

Repository contents

src/aigt_agent/                  Agent, online search, detector adapters, runners
data/benchmark_4k/               Four balanced 1K benchmark datasets
data/adversarial/                Original and adversarially paraphrased AI texts
experiments/                     Seven flat experiment entrypoints
results/main/                    Main MPU+Binoculars experiment and online evidence
results/detector_baselines/      Five standalone ML detectors
results/agent_variants/          Five ordered detector ensembles
results/llm_backbones/           Three LLM backbones
results/prompting_strategies/    S1-S4 prompt comparison
results/no_external_evidence/    Online-search removal ablation
results/adversarial_paraphrasing/Robustness under adversarial paraphrasing

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Set the required API keys in .env. Model weights are not included. Set the model directories shown in .env.example before running live detectors.

Third-party detector implementations are not vendored. Install their official releases separately, including:

• Binoculars
• GPTZero

MPU, ArguGPT, and RADAR checkpoints must be obtained from their original project/model releases and pointed to with MPU_MODEL_PATH, ARGUGPT_MODEL_PATH, and RADAR_MODEL_PATH. Gated checkpoints additionally need HF_TOKEN.

Running experiments

python experiments/main.py \
  -d data/benchmark_4k/federal_1k.csv \
  --parallel