Post
2895
10 Latest Preference Optimization Techniques
Models need feedback on what makes outputs “good” or “bad.” Policy optimization (PO) turns preferences and rewards into actual training signals. This field is evolving quickly, moving far beyond classics like PPO and GRPO. So here is our overview of 10 newest PO methods:
1. Pref-GRPO → Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning (2508.20751)
Stabilizes text-to-image reinforcement learning (RL) with pairwise preference rewards and a unified UNIGENBENCH benchmark
2. PVPO (Policy with Value Preference Optimization) → PVPO: Pre-Estimated Value-Based Policy Optimization for Agentic Reasoning (2508.21104)
This critic-free RL method uses a pre-trained model as a reference anchor to reduce bias and guide learning, selecting high-value examples through data pre-sampling
3. DCPO (Dynamic Clipping Policy Optimization) → DCPO: Dynamic Clipping Policy Optimization (2509.02333)
Uses dynamic clipping, which adjusts probability limits per token for better token exploration, and smooth reward standardization to balance rewards over training steps and prevent wasted updates
4. ARPO (Agentic Reinforced Policy Optimization) → Agentic Reinforced Policy Optimization (2507.19849)
Optimizes multi-turn LLM agents that use external tools. It uses an entropy-based adaptive rollout to explore post-tool use and an advantage attribution method to better assign credit across steps, leading to more efficient tool use with fewer resources
5. GRPO-RoC (Group Relative Policy Optimization with Resampling-on-Correct) → rStar2-Agent: Agentic Reasoning Technical Report (2508.20722)
Oversamples rollouts, then resamples them to keep diverse mistakes and only the highest-quality correct answers. It reduces noises and ends up with stronger reasoning in a code environment
Read further below ⬇️
If you like this, also subscribe to the Turing post: https://www.turingpost.com/subscribe
Models need feedback on what makes outputs “good” or “bad.” Policy optimization (PO) turns preferences and rewards into actual training signals. This field is evolving quickly, moving far beyond classics like PPO and GRPO. So here is our overview of 10 newest PO methods:
1. Pref-GRPO → Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning (2508.20751)
Stabilizes text-to-image reinforcement learning (RL) with pairwise preference rewards and a unified UNIGENBENCH benchmark
2. PVPO (Policy with Value Preference Optimization) → PVPO: Pre-Estimated Value-Based Policy Optimization for Agentic Reasoning (2508.21104)
This critic-free RL method uses a pre-trained model as a reference anchor to reduce bias and guide learning, selecting high-value examples through data pre-sampling
3. DCPO (Dynamic Clipping Policy Optimization) → DCPO: Dynamic Clipping Policy Optimization (2509.02333)
Uses dynamic clipping, which adjusts probability limits per token for better token exploration, and smooth reward standardization to balance rewards over training steps and prevent wasted updates
4. ARPO (Agentic Reinforced Policy Optimization) → Agentic Reinforced Policy Optimization (2507.19849)
Optimizes multi-turn LLM agents that use external tools. It uses an entropy-based adaptive rollout to explore post-tool use and an advantage attribution method to better assign credit across steps, leading to more efficient tool use with fewer resources
5. GRPO-RoC (Group Relative Policy Optimization with Resampling-on-Correct) → rStar2-Agent: Agentic Reasoning Technical Report (2508.20722)
Oversamples rollouts, then resamples them to keep diverse mistakes and only the highest-quality correct answers. It reduces noises and ends up with stronger reasoning in a code environment
Read further below ⬇️
If you like this, also subscribe to the Turing post: https://www.turingpost.com/subscribe