Research map / Reinforcement learning / RL from feedback
PS-PPO : Prefix-Sampling PPO for Critic-Free RLHF
Doo Hwan Hwang, Kee-Eung Kim
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 |
|---|---|---|---|
| optimize the monotone cutoff distribution instead of uniform (ablation) | PS-PPO, Qwen2.5-Math-7B · AIME25 | pass@1 accuracy (%): 9.0 -> 13.3 | Improved |
| optimize cutoff distribution instead of heuristic post-processing (ablation) | PS-PPO, Qwen2.5-Math-7B · AIME25 | pass@1 accuracy (%): 10.0 -> 13.3 | Improved |
| optimize cutoff instead of fixed-length prefix (ablation) | PS-PPO, Qwen2.5-Math-7B · AIME24 | pass@1 accuracy (%): 10.0 -> 16.7 | Improved |
| use exponential time prior instead of uniform cutoff (ablation) | PS-PPO · math benchmarks | pass@1 accuracy (%) | No clear effect |
| raise budget B from 64 to 128 (ablation) | PS-PPO, Qwen2.5-Math-7B · six math benchmarks average | average pass@1 (%): 34.0 -> 37.9 | Improved |
| raise budget B from 128 to 512 (ablation) | PS-PPO, Qwen2.5-Math-7B · six math benchmarks average | average pass@1 (%): 37.9 -> 37.8 | No clear effect |
| raise budget B from 128 to 512: time (ablation) | PS-PPO · training time per step | seconds: 1.77 -> 2.56 | Hurt |
| raise rollouts K from 2 to 16 (ablation) | PS-PPO, Qwen2.5-Math-7B · six math benchmarks average | average pass@1 (%): 34.2 -> 38.5 | Improved |
| raise rollouts K from 8 to 16 (ablation) | PS-PPO · six math benchmarks average | average pass@1 (%): 37.9 -> 38.5 | No clear effect |
| use PS-PPO prefix truncation instead of DAPO full-sequence | Qwen2.5-Math-7B · training time per step | seconds: 3.23 -> 1.77 | Improved |
| PS-PPO instead of DAPO at Tmax=4096 | Qwen2.5-Math-7B · update-stage time | speedup: 3.3x speedup over DAPO | Improved |
| PS-PPO instead of DAPO at Tmax=2048 | Qwen2.5-Math-7B · average of MATH500, AIME24, AIME25 | pass@1 accuracy (%): 42.1 -> 42.1 | No clear effect |
Closest papers by meaning
- DPO Meets PPO: Reinforced Token Optimization for RLHF (ICML 2025 · similarity 0.90)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.89)
- Policy Filtration for RLHF to Mitigate Noise in Reward Models (ICML 2025 · similarity 0.89)
- Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.89)
- ReMax: A Simple, Effective, and Efficient Reinforcement Learning Method for Aligning Large Language Models (ICML 2024 · similarity 0.89)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.88)
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.
