Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Minimizing Mismatch Risk: A Prototype-Based Routing Framework for Zero-shot LLM-generated Text Detection
Ke Sun, Guangsheng Bao, Han Cui, Yue Zhang
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 |
|---|---|---|---|
| route each input to a surrogate via DetectRouter instead of fixed LLaMA3-8B | Fast-DetectGPT · EvoBench | AUROC (%): 81.01 -> 90.85 | Improved |
| route each input via DetectRouter instead of fixed surrogate | Fast-DetectGPT · MAGE | AUROC (%): 69.69 -> 77.92 | Improved |
| apply routing to Entropy criterion | Entropy detector · EvoBench | AUROC (%): 37.46 -> 89.70 | Improved |
| apply routing to Likelihood criterion | Likelihood detector · EvoBench | AUROC (%): 77.4 -> 81.6 | Improved |
| add Stage 2 distributional alignment to Stage 1 prototypes (ablation) | DetectRouter · EvoBench | AUROC (%): 72.87 -> 90.85 | Improved |
| add Stage 1 prototype construction before Stage 2 (ablation) | DetectRouter · EvoBench | AUROC (%): 87.39 -> 90.85 | Improved |
| use KL divergence instead of L1 regression (ablation) | DetectRouter Stage 2 · EvoBench | AUROC (%): 87.07 -> 90.85 | Improved |
| add margin separation loss Lsep (ablation) | DetectRouter · MAGE | AUROC (%): 74.41 -> 77.92 | Improved |
| add prototype anchoring loss (ablation) | DetectRouter Stage 2 · EvoBench | AUROC (%): 89.56 -> 90.85 | Improved |
| use prototype metric-space routing instead of direct classification (ablation) | DetectRouter · MAGE | AUROC (%): 75.34 -> 77.92 | Improved |
| use LLM2Vec-LLaMA3 encoder instead of RoBERTa-base (ablation) | DetectRouter · EvoBench | AUROC (%): 88.34 -> 90.85 | Improved |
| take max score over surrogate pool instead of routing (ablation) | zero-shot detectors · EvoBench | AUROC (%): 81.01 -> 72.33 | Hurt |
Closest papers by meaning
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.90)
- Zero-Shot Detection of LLM-Generated Text via Implicit Reward Model (NeurIPS 2025 · similarity 0.90)
- Glimpse: Enabling White-Box Methods to Use Proprietary Models for Zero-Shot LLM-Generated Text Detection (ICLR 2025 · similarity 0.89)
- Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024 · similarity 0.88)
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.88)
- DPIC: Decoupling Prompt and Intrinsic Characteristics for LLM Generated Text Detection (NeurIPS 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.
