Glimpse

This code is for our ICLR 2025 paper "Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection", where we borrow some code from Fast-DetectGPT.

Paper | LocalDemo | OnlineDemo | OpenReview

• 9/4/2025: The demo server is facing a hardware problem, and we are actively working on a solution.
• 🔥 [4/14/2025] Free API access to the detectors is ready!
• 🔥 [2/10/2025] Local and online demos are ready! As OpenAI and AzureOpenAI have discontinued the legacy Completion API for GPT-3.5-turbo and GPT-4, our demo is limited to davinci and babbage models.

Brief Intro

Glimpse serves as a bridge between white-box methods, which rely on local LLMs for scoring, and proprietary LLMs. It estimates full distributions based on partial observations from API-based models. Our empirical analysis demonstrates that detection methods leveraging these estimated distributions achieve detection accuracies comparable to those obtained using real distributions.

The main results are as follows, where Fast-DetectGPT using the proprietary GPT-3.5 achieves significantly better detection accuracies than that using open-source GPT-Neo-2.7B, especially on multiple languages.

Local Demo

Run following command locally for an interactive demo:

python scripts/local_infer.py  --api_key <openai API key>  --scoring_model_name davinci-002 

An example looks like

Please enter your text: (Press Enter twice to start processing)
工作量和工作强度会根据银行的不同而有所不同。但一般来说，作为业务员需要在工作中需要面对各类客户，以及承担一定的工作压力和业绩指标，因此这个职业确实需要相当的努力和不断的自我提高。

Glimpse criterion is -0.3602, suggesting that the text has a probability of 69% to be machine-generated.

Environment

• Python3.12
• Setup the environment: pip install -r requirements.txt

(Notes: the baseline methods are run on 1 GPU of Tesla A100 with 80G memory, while Glimpse is run on a CPU environment.)

Experiments

Following folders are created for our experiments:

• ./exp_main -> experiments with five latest LLMs as the source model (main.sh).
• ./exp_langs -> experiments on six languages (langs.sh).

(Notes: we share the data and results for convenient reproduction.)

Citation

If you find this work useful, you can cite it with the following BibTex entry:

@articles{bao2025glimpse,
  title={Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection},
  author={Bao, Guangsheng and Zhao, Yanbin and He, Juncai and Zhang, Yue},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}