Research map / Large language models / Alignment and preferences
Distributionally Robust Reinforcement Learning from Human Feedback
Debmalya Mandal, Paulius Sasnauskas, Goran Radanovic
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 with distributionally robust loss (rho=0.2) | Bradley-Terry reward model · RewardBench (reward model, 400K Unified-Feedback) | Avg accuracy: 72.3 -> 72.9 | Improved |
| train reward model with distributionally robust loss (rho=0.2) | Bradley-Terry reward model · RewardBench (reward model, 400K Unified-Feedback) | Reasoning accuracy: 66.8 -> 71.4 | Improved |
| increase robustness radius rho to 1.0 for reward model (ablation) | DR reward model · RewardBench (reward model, 400K Unified-Feedback) | Avg accuracy: 72.3 -> 46.4 | Hurt |
| increase robustness radius rho from 0.2 to 0.4 for reward model (ablation) | DR reward model · RewardBench (reward model, 400K Unified-Feedback) | Avg accuracy: 72.9 -> 61.0 | Hurt |
| use DR PPO with rho=0.1 instead of standard PPO | PPO with BTL reward model · RewardBench (PPO policy, 5K Unified-Feedback) | Reasoning accuracy: 45.9 -> 55.9 | Improved |
| use DR PPO with rho=0.1 instead of standard PPO | PPO with BTL reward model · RewardBench (PPO policy, 5K Unified-Feedback) | Avg accuracy: 48.7 -> 52.3 | Improved |
| use DR PPO rho=0.6 instead of standard PPO (ablation) | PPO · RewardBench (PPO policy, 5K Unified-Feedback) | Avg accuracy: 48.7 -> 48.8 | No clear effect |
| use DR PPO rho=0.1 instead of standard PPO | PPO · HHH Alignment (PPO) | accuracy: 66.9 ± 0.5 -> 68.1 ± 0.3 | Improved |
| use DR PPO rho=0.1 instead of standard PPO | PPO · MT Bench (PPO) | accuracy: 59.9 ± 0.3 -> 61.6 ± 1.0 | Improved |
| use DR PPO rho=0.1 instead of standard PPO | PPO · Unified-Feedback test (PPO) | accuracy: 59.8 ± 0.2 -> 59.8 ± 0.4 | No clear effect |
| use DR DPO rho=0.8 instead of standard DPO | DPO · Unified-Feedback test (DPO) | accuracy: 60.9 ± 0.02 -> 64.4 ± 0.3 | Improved |
| use DR DPO rho=0.8 instead of standard DPO | DPO · HHH Alignment (DPO) | accuracy: 69.2 ± 0.1 -> 70.6 ± 0.4 | Improved |
Closest papers by meaning
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.92)
- Robust Reinforcement Learning from Corrupted Human Feedback (NeurIPS 2024 · similarity 0.92)
- Distributionally Robust Token Optimization in RLHF (NeurIPS 2026 · similarity 0.92)
- Adaptive Preference Scaling for Reinforcement Learning with Human Feedback (NeurIPS 2024 · similarity 0.91)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Safe RLHF: Safe Reinforcement Learning from Human Feedback (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.
