Research map / Large language models / Alignment and preferences
Beyond Binary Preferences: A Principled Framework for Reward Modeling with Ordinal Feedback
Amirhossein Afsharrad, Ruida Zhou, Luca Viano, Sanjay Lall, Mohammad Ghavamzadeh
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 |
|---|---|---|---|
| Use NLL-Symmetric ordinal loss instead of Simple BT | reward model, Llama-3.1-8B · RewardBench (HelpSteer2) | number of errors: 433 -> 282 | Improved |
| Use NLL-Symmetric ordinal loss instead of Simple BT | reward model, Llama-3.1-8B · RewardBench (HelpSteer2) | mean error margin: 3.827 -> 0.501 | Improved |
| Train thresholds jointly instead of post-hoc calibration of Soft Label (ablation) | ordinal reward model, Llama-3.1-8B · HelpSteer2 test | MAE: 1.725 -> 1.060 | Improved |
| Train thresholds jointly instead of post-hoc calibration of Margin BT (ablation) | ordinal reward model, Llama-3.1-8B · HelpSteer2 test | Acc@0: 16.1% -> 29.7% | Improved |
| Use symmetric thresholds (NLL-Sym) instead of asymmetric (NLL-Asym) (ablation) | NLL ordinal loss, Llama-3.1-8B · RewardBench Avg (HelpSteer2) | accuracy: 0.809 -> 0.843 | Improved |
| Use symmetric thresholds (NLL-Sym) instead of asymmetric (NLL-Asym) (ablation) | NLL ordinal loss, Llama-3.1-8B · RM-Bench Total (HelpSteer2) | accuracy: 0.616 -> 0.651 | Improved |
| Use NLL-Sym instead of Margin BT | reward model, Llama-3.1-8B · RewardBench Avg (HelpSteer2) | accuracy: 0.802 -> 0.843 | Improved |
| Use NLL-Sym instead of Margin BT | reward model, Mistral-7B · RewardBench Avg (HelpSteer2) | accuracy: 0.749 -> 0.779 | Improved |
| Use All-Threshold instead of Margin BT | reward model, Mistral-7B · RM-Bench Total (HelpSteer3) | accuracy: 0.620 -> 0.565 | Hurt |
| Use NLL-Sym instead of Scaled BT | reward model, Llama-3.1-8B · RM-Bench Total (HelpSteer3) | accuracy: 0.673 -> 0.676 | No clear effect |
| Corrupt 100% of labels with systematic shift noise (ablation) | NLL-Symmetric · RewardBench Avg | accuracy: 0.843 -> 0.808-0.846 | No clear effect |
| Add L2 regularization on thresholds (ablation) | ordinal losses · threshold convergence | threshold stability | Improved |
Closest papers by meaning
- Reward Modeling with Ordinal Feedback: Wisdom of the Crowd (ICML 2025 · similarity 0.92)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.91)
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.91)
- Transforming and Combining Rewards for Aligning Large Language Models (ICML 2024 · similarity 0.91)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.91)
- Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment (ICML 2025 · 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.
