Research map / Large language models / Language, safety and interpretability
DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeTeCtive, which fine-tunes text encoders with a multi-task multi-level contrastive loss plus a KNN retrieval pipeline to detect AI-generated text, and adds training-free incremental adaptation via database augmentation.
Problem
Detecting AI-generated text with supervised binary classifiers generalizes poorly to out-of-distribution domains and newly emerged LLMs.
Approach
Reframe detection as distinguishing writing styles of different sources (human vs. LLM families, companies, and individual models) using a multi-level contrastive loss with hierarchical positive/negative sets, combined with a KNN retrieval over a feature database, so new OOD data can be added to the database without retraining.
Key results
- On Deepfake Cross-domains & Cross-models, DeTeCtive reaches 96.15 AvgRec, versus 93.49 for T5-Sentinel and 90.53 for Longformer.
- On Deepfake Unseen Domains, AvgRec rises from 82.60 to 89.63 with TFIA (database augmentation), and Unseen Models from 92.19 to 93.03.
- On M4-monolingual, DeTeCtive reaches 98.44 AvgRec versus 91.92 for the second-best (SCL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeTeCtive instead of the strongest reported baseline | AI-generated text detection · M4-monolingual | AvgRec: 91.92 -> 98.44 | Helped |
| Use DeTeCtive instead of the strongest reported baseline | AI-generated text detection · M4-multilingual | AvgRec: 86.27 -> 93.42 | Helped |
| Use DeTeCtive instead of the strongest reported baseline | AI-generated text detection · TuringBench | AvgRec: 99.59 -> 99.74 | Helped |
| Use DeTeCtive instead of the strongest reported baseline | AI-generated text detection · Deepfake Cross-domains & Cross-models | AvgRec: 93.49 -> 96.15 | Helped |
| Add Cross-domains & Cross-models data to the feature database | DeTeCtive · Deepfake Unseen Models | AvgRec: 92.19 -> 93.03 | Helped |
| Add Cross-domains & Cross-models data to the feature database | DeTeCtive · Deepfake Unseen Domains | AvgRec: 82.60 -> 89.63 | Helped |
| Use multi-level contrastive loss instead of plain contrastive loss (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 94.22 -> 96.15 | Helped |
| Add the auxiliary cross-entropy loss (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 95.01 -> 96.15 | Helped |
| Use the alpha-weighted loss term (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 95.25 -> 96.15 | Helped |
| Use the beta-weighted loss term (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 94.94 -> 96.15 | Helped |
| Use the gamma-weighted loss term (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 94.88 -> 96.15 | Helped |
| Use KNN instead of the classification head (ablation) | DeTeCtive · Deepfake Cross-domains & Cross-models | AvgRec: 93.19 -> 96.15 | Helped |
Surprising findings
- Using a standard classification head instead of KNN lowers AvgRec from 96.15 to 93.19 on the Deepfake subset, showing the retrieval-based classification matters.
- Binoculars achieves only 51.24 AvgRec on TuringBench, near chance, despite strong results elsewhere.
Limitations
- Interpretability of the learned writing-style features is not explored.
- Training on a larger corpus was not carried out.
- Model-family hierarchy assumptions (same-company relations) are not validated beyond the chosen datasets.
Experimental setup
Datasets: Deepfake M4 TuringBench OUTFOX DIPPER
Benchmarks: Deepfake M4-monolingual M4-multilingual TuringBench OUTFOX
Metrics: AvgRec F1 HumanRec MachineRec
Baselines: RoBERTa SCL Longformer T5-Sentinel Binoculars FastText GLTR DetectGPT OUTFOX HC3 Detector Random Forest SVM
Builds on: SimCLR SimCSE supervised contrastive learning dense information retrieval Faiss
Closest papers by meaning
- Position: On the Possibilities of AI-Generated Text Detection (ICML 2024 · similarity 0.89)
- Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection (NeurIPS 2025 · similarity 0.88)
- Raidar: geneRative AI Detection viA Rewriting (ICLR 2024 · similarity 0.88)
- DETree: DEtecting Human-AI Collaborative Texts via Tree-Structured Hierarchical Representation Learning (NeurIPS 2025 · similarity 0.88)
- Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection (AAAI 2025 · similarity 0.87)
- Towards Automatic Boundary Detection for Human-AI Collaborative Hybrid Essay in Education (AAAI 2024 · 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.