Research map / Large language models / Alignment and preferences
RewardEval: Advancing Reward Model Evaluation
Saumya Malik, Valentina Pyatkin, Sander Land, Jacob Morrison, Noah A. Smith, Hannaneh Hajishirzi, Nathan Lambert
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 |
|---|---|---|---|
| train reward model on combined Tulu and Skywork preference data instead of either alone (ablation) | Tulu 8B SFT reward model · RewardBench 2 | average accuracy: 63.5 (Tulu) and 66.7 (Skywork) -> 68.2 (both) | Improved |
| initialize reward model from Llama 8B Instruct instead of Llama 8B Base (ablation) | Llama 8B reward model · RewardBench 2 | average accuracy: 64.9 -> 72.8 | Improved |
| initialize reward model from Qwen 7B Instruct instead of Qwen 7B Base (ablation) | Qwen 7B reward model · RewardBench 2 | average accuracy: 68.2 -> 73.3 | Improved |
| initialize from Tulu DPO or RL rather than Tulu SFT (ablation) | Tulu 8B reward model · RewardBench 2 | average accuracy: 68.2 -> 68.7; Safety 89.8 -> 86.0 | Mixed |
| add Skywork preference data to Tulu data (ablation) | Tulu 8B SFT reward model · RewardBench 2 Focus | accuracy: 71.3 -> 93.7 Focus; 81.1 -> 88.0 Safety | Improved |
| use Qwen 7B Instruct as reward model base for Math (ablation) | reward model · RewardBench 2 Math | accuracy: Qwen 7B Instruct 71.6 vs Tulu 70B SFT 67.8 | Improved |
| train reward models for two epochs rather than one (ablation) | reward model training · RewardBench 2 | average accuracy: eight among the eighteen best models trained for two epochs | Improved |
| scale reward model base from 8B to 70B (ablation) | Tulu SFT reward model, both data · RewardBench 2 | average accuracy: 68.2 -> 72.2 | Improved |
| use benchmark average to predict best-of-N downstream performance | RewardBench 2 · BoN over 113 RMs | Pearson correlation: 0.87 | Improved |
| use reward model from different lineage than the policy in PPO (ablation) | PPO with Tulu 3 8B SFT policy · Tulu 3 eval suite average | average score: 72.9 RB2 -> 54.5 PPO versus 49.8 RB2 -> 60.2 PPO for on-policy RM | Hurt |
| train PPO with the best on-policy reward model instead of the SFT policy or DPO model | Tulu 3 8B SFT · Tulu 3 eval suite average | average score: 54.1 -> 60.3 | Improved |
| increase reward model RB2 score among decent on-policy RMs for PPO | PPO · Tulu 3 eval suite average | average score: RB2 49.8 to 68.5 all give similar PPO score | No clear effect |
Closest papers by meaning
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.94)
- How to Evaluate Reward Models for RLHF (ICLR 2025 · similarity 0.93)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.93)
- Post-hoc Reward Calibration: A Case Study on Length Bias (ICLR 2025 · similarity 0.92)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.92)
- Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree? (ICLR 2025 · similarity 0.92)
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.
