Research map / Large language models / Language, safety and interpretability
HLD: Approximate Hierarchical Linguistic Distribution Modeling for LLM-Generated Text Detection
Rui Guo, Weibin Zeng, Fuzhang Wu, Yan Kong, sicheng shen, Yanjun Wu, Weiming Dong
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 word-level features (ablation) | HLD · DetectRL Review (train on Arxiv) | AUROC: 84.43 -> 91.54 | Improved |
| add POS features (ablation) | HLD · DetectRL Review (train on Arxiv) | AUROC: 88.10 -> 91.54 | Improved |
| add dependency features (ablation) | HLD · DetectRL Review (train on Arxiv) | AUROC: 90.05 -> 91.54 | Improved |
| add semantic features (ablation) | HLD · DetectRL Review (train on Arxiv) | AUROC: 87.31 -> 91.54 | Improved |
| add semantic features (ablation) | HLD · DetectRL Writing (train on Arxiv) | AUROC: 76.24 -> 79.40 | Improved |
| add POS features (ablation) | HLD · DetectRL Writing F1 (train on Arxiv) | F1: 72.99 -> 71.90 | Hurt |
| use HLD instead of RoBERTa-base | HLD vs Roberta-base · DetectRL multi-LLM avg | AUROC: 98.24 -> 99.12 | Improved |
| use HLD instead of RoBERTa-base | HLD vs Roberta-base · DetectRL cross-domain (train on Arxiv) | AUROC: 81.38 -> 86.61 | Improved |
| use HLD instead of RoBERTa-base | HLD vs Roberta-base · DetectRL paraphrase attack | AUROC: 73.00 -> 97.04 | Improved |
| use HLD instead of DPIC | HLD vs DPIC · DetectRL adversarial attacks avg | AUROC: 93.03 -> 97.85 | Improved |
| increase n-gram context length k beyond 5 | HLD word-level · DetectRL | AUROC | Hurt |
| use HLD instead of Fast-DetectGPT/RAIDAR proxy models | HLD · inference time | total time (s): 35.13 -> 5.24 | Improved |
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.
- RADAR — RADAR: Robust AI-Text Detection via Adversarial Learning (NeurIPS 2023) · better on 2 of 2 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
- Training-free LLM-generated Text Detection by Mining Token Probability Sequences (ICLR 2025 · similarity 0.91)
- 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 0.90)
- Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection (NeurIPS 2025 · similarity 0.90)
- Position: On the Possibilities of AI-Generated Text Detection (ICML 2024 · similarity 0.90)
- Multiscale Positive-Unlabeled Detection of AI-Generated Texts (ICLR 2024 · similarity 0.89)
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.
