Research map / Large language models / Language, safety and interpretability
Intrinsic Dimension Estimation for Robust Detection of AI-Generated Texts
Eduard Tulchinskii, Kristian Kuznetsov, Kushnareva Laida, Daniil Cherniavskii, Sergey Nikolenko, Evgeny Burnaev, Serguei Barannikov, Irina Piontkovskaya
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 |
|---|---|---|---|
| Use PHD instead of general-purpose detector baselines | Artificial text detection · GPT-3.5 generations from Wikipedia prompts | accuracy at 1% FPR: 46.7 vs 40.0 | Hurt |
| Apply DIPPER paraphrasing to generated text | PHD detector · GPT-3.5 generations from Wikipedia prompts | accuracy at 1% FPR: 40.0 -> 41.2 | Improved |
| Apply DIPPER paraphrasing to generated text | MLE detector · GPT-3.5 generations from Wikipedia prompts | accuracy at 1% FPR: 46.7 -> 33.3 | Hurt |
| Transfer a supervised RoBERTa classifier across domains instead of using PHD | Artificial text detection · Wikipedia-to-Reddit | accuracy: 0.781 vs 0.535 | Hurt |
| Use PHD rather than the supervised RoBERTa classifier for cross-model transfer | Artificial text detection · GPT2-to-OPT | accuracy: 0.759 vs 0.993 | Hurt |
| Use PHD rather than the supervised RoBERTa classifier in-domain | Artificial text detection · GPT3.5-to-GPT3.5 | accuracy: 0.843 vs 0.990 | Hurt |
| Apply DIPPER lexical diversity and reordering (Lex 60, Order 60) | PHD measurement of GPT-3.5 text · DIPPER-paraphrased GPT-3.5 generations | mean intrinsic dimension: 7.30 -> 7.51 | No clear effect |
| Apply DIPPER lexical diversity and reordering (Lex 60, Order 60) | MLE measurement of GPT-3.5 text · DIPPER-paraphrased GPT-3.5 generations | mean intrinsic dimension: 9.68 -> 10.01 | Mixed |
| Increase DIPPER lexical diversity from original to Lex 60 | PHD measurement of GPT-3.5 text · DIPPER paraphrasing | mean intrinsic dimension: 7.30 -> 7.51 | No clear effect |
| Increase DIPPER reordering to Order 60 | PHD measurement of GPT-3.5 text · DIPPER paraphrasing | mean intrinsic dimension: 7.42 -> 7.45 | No clear effect |
Closest papers by meaning
- Position: On the Possibilities of AI-Generated Text Detection (ICML 2024 · similarity 0.90)
- Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection (NeurIPS 2025 · similarity 0.87)
- AI as Humanity’s Salieri: Quantifying Linguistic Creativity of Language Models via Systematic Attribution of Machine Text against Web Text (ICLR 2025 · similarity 0.87)
- Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text (NeurIPS 2026 · similarity 0.87)
- DETree: DEtecting Human-AI Collaborative Texts via Tree-Structured Hierarchical Representation Learning (NeurIPS 2025 · similarity 0.86)
- Towards Automatic Boundary Detection for Human-AI Collaborative Hybrid Essay in Education (AAAI 2024 · similarity 0.86)
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.
