Research map / Large language models / Language, safety and interpretability
D&R: Recovery-based AI-Generated Text Detection via a Single Black-box LLM Call
Yuxia Sun, Ran Zhang, Aoxiang Sun, Xu Li, Zitao Liu, Jingcai Guo
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 |
|---|---|---|---|
| add D&R single-call recovery detection instead of RAIDAR multi-call rewriting | RAIDAR · Long-text datasets avg | AUROC: 0.8770 -> 0.9602 | Improved |
| use D&R instead of RAIDAR on short texts | RAIDAR · Short-text datasets | AUROC: 0.7248 -> 0.8687 | Improved |
| use Within-Chunk Shuffling instead of Global Shuffling (ablation) | D&R · Long-text avg | AUROC: 0.5564 -> 0.9602 | Improved |
| use Within-Chunk Shuffling instead of Chunk-Order Shuffling (ablation) | D&R · Long-text avg | AUROC: 0.7423 -> 0.9602 | Improved |
| add semantic similarity (BERTScore F1) to the classifier features (ablation) | D&R · Long-text avg, 4 advanced source models | AUROC: 0.6906 -> 0.9614 | Improved |
| add structural similarity (Kendall tau, Spearman rho) to the classifier features (ablation) | D&R · Long-text avg, 4 advanced source models | AUROC: 0.7707 -> 0.9614 | Improved |
| use a small local recovery model (Mistral-7B) instead of DeepSeek-v3 (ablation) | D&R · Long-text, 4 advanced source models | AUROC: 0.9614 -> 0.9359 | Hurt |
| use a different source model than the recovery model (ablation) | D&R · Long-text avg | AUROC: 0.9793 -> 0.9603 | Hurt |
| use a different source model than the recovery model (ablation) | RAIDAR · Long-text avg | AUROC: 0.9712 -> 0.8792 | Hurt |
| apply paraphrase attack to AI text (ablation) | D&R · RAID benchmark | AUROC: 0.8736 -> 0.8210 | Hurt |
| replace multi-call RAIDAR by single-call D&R for efficiency | RAIDAR · per-sample latency | seconds: 15 -> 2 | Improved |
| increase recovery-model temperature from 0.2 to 0.8 (ablation) | D&R recovery similarity · ML-ArXiv-Papers | recovery similarity | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Fast-DetectGPT — Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024) · better on 2 of 2 shared results
- DeTeCtive — DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning (NeurIPS 2024) · better on 1 of 1 shared result
- Dald — DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024) · better on 1 of 1 shared result
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 0.89)
- Position: On the Possibilities of AI-Generated Text Detection (ICML 2024 · similarity 0.89)
- Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection (ICLR 2025 · similarity 0.88)
- DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm (NeurIPS 2025 · similarity 0.88)
- Raidar: geneRative AI Detection viA Rewriting (ICLR 2024 · similarity 0.88)
- DPIC: Decoupling Prompt and Intrinsic Characteristics for LLM Generated Text Detection (NeurIPS 2024 · similarity 0.88)
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.
