Research map / Large language models / Language, safety and interpretability
Black-Box Detection of LLM-Generated Text Using Generalized Jensen Shannon Divergence
Shuangyi Chen, Ashish J Khisti
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| summarize surprisal dynamics with a GJS gap against human and machine references (SurpMark) | proxy-LM black-box detection · closed-source generators (GPT-5-Chat, GPT-4.1-mini, Gemini-1.5-Flash) | AUROC (avg): 69.55 (best baseline R-Detect) -> 78.32 | Improved |
| SurpMark vs best baseline on open-source generators | proxy-LM black-box detection · 9 open-source generators | AUROC (avg): 90.04 (Lastde++) -> 91.23 | Improved |
| use k-means instead of equal-width binning (ablation) | SurpMark discretization · WritingPrompts@GPT5-chat (k=7) | AUROC: 62.90 -> 82.05 | Improved |
| use k-means instead of equal-mass binning (ablation) | SurpMark discretization · XSum@GPT-4.1-mini (k=7) | AUROC: 71.40 -> 80.42 | Improved |
| use k-means instead of equal-mass binning (ablation) | SurpMark discretization · WritingPrompts@GPT5-chat (k=7) | AUROC: 84.42 (equal-mass) vs 82.05 (k-means) | No clear effect |
| increase number of bins k beyond 6-7 (ablation) | SurpMark · GPT-J-6B and Llama-3.2-3B | AUROC | No clear effect |
| use higher-order Markov chain instead of first-order (ablation) | SurpMark · SQuAD and appendix datasets | AUROC | Hurt |
| second-order instead of first-order surprisal-state model (ablation) | discretized surprisal states · GPT-5-chat text | perplexity reduction: +0.528% relative perplexity gain | No clear effect |
| increase number of reference samples up to 100 (ablation) | SurpMark · reference size sweep | AUROC | Improved |
| increase test text length from 50 to about 150-200 tokens (ablation) | SurpMark · test length sweep | AUROC | Improved |
| use references from a different domain (ablation) | SurpMark · DetectRL Yelp | AUROC: 90.18 -> 85.48 (arXiv-ref) | Hurt |
| use a stronger proxy LM (ablation) | SurpMark · proxy model sweep | AUROC | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Fast-DetectGPT — Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.91)
- Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection (ICLR 2025 · similarity 0.89)
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.89)
- DPIC: Decoupling Prompt and Intrinsic Characteristics for LLM Generated Text Detection (NeurIPS 2024 · similarity 0.89)
- Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text (NeurIPS 2026 · similarity 0.87)
- AdaDetectGPT: Adaptive Detection of LLM-Generated Text with Statistical Guarantees (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
