Research map / Large language models / Language, safety and interpretability
HLPD: Aligning LLMs to Human Language Preference for Machine-Revised Text Detection
Fangqi Dai, Xingjian Jiang, Zizhuang Deng
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 |
|---|---|---|---|
| align scoring model to human style with HLPO instead of machine style (ImBD) for rewrite detection | HLPD · Rewrite detection, GPT-3.5 + GPT-4o | AUROC (overall avg): 0.7870 -> 0.9144 | Improved |
| human-aligned scoring model for advanced LLM generations | HLPD · XSum advanced LLM generation | AUROC (avg): 0.9228 -> 0.9738 | Improved |
| human-aligned scoring model under adversarial multi-task revision (GPT-4o) | HLPD · Adversarial multi-task revision, GPT-4o | AUROC (avg): 0.8441 -> 0.9464 | Improved |
| add preference alignment (HLPO) versus no alignment (ablation) | GPT-Neo-2.7B scorer · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.35 -> 0.88 | Improved |
| replace SFT/RLHF/ORPO/IPO with HLPO-style alignment (ablation) | HLPD · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.71 -> 0.94 | Improved |
| add dynamic beta to HLPO (ablation) | HLPO · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.88 -> 0.89 | Improved |
| replace sigmoid with linear contrastive loss in HLPO (ablation) | HLPO · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.88 -> 0.91 | Improved |
| combine linear loss and dynamic beta (ablation) | HLPO · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.88 -> 0.94 | Improved |
| train ImBD with the linear loss and dynamic beta (ablation) | ImBD · Adversarial multi-task revision, DeepSeek-R1 + Gemini-2.5-Pro | AUROC (avg): 0.81 -> 0.86 | Improved |
| human-aligned scoring on polish task for GPT-4o PubMed | HLPD · Rewrite detection, GPT-4o PubMed | AUROC: 0.7393 vs Ghostbuster 0.7853 | Mixed |
| increase number of revision iterations from 1 to 5 | HLPD · Multi-iterative revision, XSum | AUROC: 0.9523 -> 0.9905 | No clear effect |
| apply perturb-and-select humanization using the HLPO scorer, 4 iterations, against detection | GPT-3.5 polished XSum text · GPTZero | AI probability: 74% decrease | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ReMoDetect — ReMoDetect: Reward Models Recognize Aligned LLM's Generations (NeurIPS 2024) · better on 4 of 4 shared results
- Fast-DetectGPT — Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature (ICLR 2024) · better on 3 of 3 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
- ReMoDetect: Reward Models Recognize Aligned LLM's Generations (NeurIPS 2024 · similarity 0.91)
- Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection (AAAI 2025 · similarity 0.90)
- Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy (NeurIPS 2026 · similarity 0.88)
- Detecting Machine-Generated Texts by Multi-Population Aware Optimization for Maximum Mean Discrepancy (ICLR 2024 · similarity 0.87)
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.87)
- DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios (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.
