Research map / Large language models / Alignment and preferences
Bootstrapping LLMs via Preference-Based Policy Optimization
Chen Jia
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 |
|---|---|---|---|
| PbPO online bootstrapping (300 episodes) instead of offline DPO | LLaMA2-7B · 5-benchmark average | accuracy: 49.7 -> 57.1 (seq RM) | Improved |
| PbPO instead of Online PPO | Qwen2-7B · GSM8K | accuracy: 79.2 -> 85.2 | Improved |
| more online PbPO episodes (100 -> 300) (ablation) | PbPO seq RM, LLaMA2-7B · 5-benchmark average | accuracy: 52.7 -> 57.1 | Improved |
| token-level RM instead of sequence-level RM (ablation) | PbPO · GSM8K (LLaMA2-7B) | accuracy: 58.1 -> 60.2 | Improved |
| token-level RM instead of sequence-level RM (ablation) | PbPO · AGIEval/MMLU (LLaMA2-7B) | accuracy: AGIEval 43.7 -> 43.5; MMLU 57.1 -> 56.3 | Mixed |
| remove reward-agnostic exploration (no enhancer policy) (ablation) | PbPO · AGIEval/ARC-C/MMLU/GSM8K | accuracy | Hurt |
| remove reward-aware exploration (pipeline RM then policy) (ablation) | PbPO · AGIEval/ARC-C/MMLU/GSM8K | accuracy | Hurt |
| set conservatism beta > 0 instead of 0 (ablation) | PbPO · LLaMA2-7B | accuracy | Improved |
| larger reward model (3B up to 70B) (ablation) | PbPO · LLaMA2-7B | accuracy | Improved |
| online iterative methods (Online DPO, Online PPO, Best-of-N) vs offline RLHF | LLaMA2-7B · 5-benchmark average | accuracy: DPO 49.7 -> Online DPO 50.8 | No clear effect |
Closest papers by meaning
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 2025 · similarity 0.92)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.91)
- Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 2025 · similarity 0.91)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.90)
- Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data (ICML 2024 · similarity 0.90)
- Statistical Rejection Sampling Improves Preference Optimization (ICLR 2024 · 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.
