Research map / Large language models / Alignment and preferences
RRHF: Rank Responses to Align Language Models with Human Feedback
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ranking loss to the SFT loss (ablation) | RRHF Alpaca BP · HH dataset | average reward: -1.14 -> -1.03 | Helped |
| add sampled responses (beam/diverse beam) to dataset-provided responses (ablation) | RRHF Alpaca · HH dataset | average reward: -1.31 -> -1.02 | Helped |
| iterate sampling from updated policy (IP-1 to IP-3) (ablation) | RRHF Alpaca · HH dataset | average reward: -1.02 -> -0.94 | Helped |
| use online sampling without KL (ablation) | RRHF Alpaca · HH dataset | reward / PPL: -1.02 -> 0.34; PPL 14.75 -> 63.78 | Mixed |
| add KL penalty to online sampling (ablation) | RRHF OP-32 · HH dataset | average reward: 0.34 -> -0.86 (PPL 63.78 -> 19.76) | Helped |
| SFT warmup of initial Alpaca checkpoint (ablation) | RRHF DP · HH dataset | average reward: -1.02 -> -1.19 | Hurt |
| use diverse beam instead of vanilla beam search (ablation) | RRHF Alpaca · HH dataset | average reward: -1.03 -> -1.02 | No clear effect |
| replace PPO with RRHF | Alpaca · HH dataset | average reward: -1.03 -> -1.03 | No clear effect |
| replace PPO with RRHF | LLaMA · HH dataset | average reward: -1.62 -> -1.34 | Helped |
| replace PPO with RRHF | Alpaca-sft · HH dataset | average reward: -1.25 -> -1.19 | Helped |
| replace PPO with RRHF | GPT-2 SFT · IMDB | reward: 0.796 -> 0.861 | Helped |
Closest papers by meaning
- Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback (AAAI 2025 · similarity 0.93)
- RLTHF: Targeted Human Feedback for LLM Alignment (ICML 2025 · similarity 0.92)
- RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback (ICML 2024 · similarity 0.92)
- Preference Ranking Optimization for Human Alignment (AAAI 2024 · similarity 0.92)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.92)
- Generative RLHF-V: Learning Principles from Multi-modal Human Preference (NeurIPS 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.