Research map / Large language models / Alignment and preferences
Human-AI Curation Synergy: Scaling Preference Data Curation via Human-Guided AI Feedback
Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao, Jujie He, Jiacai Liu, Chaojie Wang, Rui Yan, Wei Shen, Fuxiang Zhang, Jiacheng Xu, Yang Liu
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 simple LLM curation of unverified pairs (ablation) | Skywork-Reward-V2 pipeline · Stage 1 iter 1, average of benchmarks | average score: 0.1 | No clear effect |
| add human curation of unverified pairs (ablation) | Skywork-Reward-V2 pipeline · Stage 1 iter 1, average of benchmarks | average score: 2.3 | Improved |
| add adaptive retrieval of gold examples to LLM annotation (ablation) | Skywork-Reward-V2 pipeline · Stage 1 iter 1, average of benchmarks | average score: 0.9 | Improved |
| use human annotators with preference attributes instead of raw annotation (ablation) | human curation · Stage 1 iter 1, average of benchmarks | average score: 71.4 -> 72.1 | Improved |
| use full verification protocol with external tools instead of attributes only (ablation) | human curation · Stage 1 iter 1, average of benchmarks | average score: 72.1 -> 74.2 | Improved |
| raw human curation instead of no curation (ablation) | human curation · Stage 1 iter 1, average of benchmarks | average score: 71.0 -> 71.4 | Improved |
| scale uncurated preference data (ablation) | SynPref-40M early version · average of six benchmarks | average score | No clear effect |
| scale curated preference data (ablation) | SynPref-40M early version · average of six benchmarks | average score | Improved |
| recycle discarded pairs by flipping chosen and rejected (ablation) | Skywork-Reward-V2 · average of six benchmarks | average score | Improved |
| train on only 1.8% (about 290K samples) of the curated mixture (ablation) | Skywork-Reward-V2-Llama-3.1-8B preliminary · average of six benchmarks | average score | Improved |
| add 14M flipped discarded pairs to the 26M curated pairs (40M variant) | Skywork-Reward-V2-Llama-3.1-8B · seven benchmarks average | average score: 85.8 -> 88.6 | Improved |
| increase backbone from Qwen3-0.6B to Qwen3-8B | Skywork-Reward-V2 · seven benchmarks average | average score: 70.9 -> 79.3 | Improved |
Closest papers by meaning
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.92)
- Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning (NeurIPS 2024 · similarity 0.92)
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.92)
- HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages (NeurIPS 2025 · similarity 0.92)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.92)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.91)
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.
