Research map / Large language models / Alignment and preferences
Probing Preference Representations: A Multi-Dimensional Evaluation and Analysis Method for Reward Models
Chenglong Wang, Yifu Huo, Yang Gan, Yongyu Mu, Qiaozhi He, Murun Yang, Bei Li, Chunliang Zhang, Tongran Liu, Anxiang Ma, Zhengtao Yu, Jingbo Zhu and 1 more
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 inference-time-probing sample filtering to PPO | LLaMA-3.1-8B-RM-400k PPO · Aligned LLaMA-3.1-8B-SFT | win rate: 57.3 -> 62.5 | Improved |
| discard random samples instead of probing-selected ones | PPO · Aligned LLaMA-3.1-8B-SFT | win rate: 57.3 -> 54.3 | Hurt |
| lower threshold d_tau from 140 to 100 (ablation) | inference-time probing · Threshold sweep | win rate: 63.5 -> 59.6 | Hurt |
| raise threshold d_tau from 140 to 180 (ablation) | inference-time probing · Threshold sweep | win rate: 63.5 -> 60.1 | Hurt |
| increase cluster size from 2 to 5 (ablation) | inference-time probing · Cluster size sweep | win rate: 63.5 -> 59.2 | Hurt |
| use inference-time probing with any threshold vs vanilla (ablation) | PPO · Threshold sweep | win rate: 57.4 -> 59.6 | Improved |
| fuse MRMBench with RM-Bench | reward model evaluation · Correlation with aligned LLM win rate | Pearson correlation: 0.89 -> 0.92 | Improved |
| compare MRMBench with RewardBench as predictor | reward model evaluation · Correlation with aligned LLM win rate | Pearson correlation: 0.34 -> 0.89 | Improved |
| use top-most layer instead of lower layers for probing (ablation) | GPM-Llama-3.1-8B and GPM-Gemma-2B · MRMBench probing tasks | accuracy | Improved |
| compare easy vs hard probing tasks | GPM-LLaMA-3.1-8B · MRMBench | average accuracy: 79.6 -> 64.5 | Hurt |
Closest papers by meaning
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.91)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.91)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.91)
- Transforming and Combining Rewards for Aligning Large Language Models (ICML 2024 · similarity 0.90)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.90)
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · similarity 0.90)
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.
