SP3O Improves RL from Segment Preferences Without Reward Models

Evan Assmus, Qining Zhang, Lei Ying· August 5, 2026 View original

Key takeaways

  • SP3O is a new reward-model-free, critic-free, gradient-based PbRL algorithm.
  • It uses segment-level preference feedback, which is easier for human evaluators.
  • SP3O constructs a policy value difference estimator via off-policy importance sampling.
  • It outperforms other PbRL/RLHF algorithms in robotic control and LLM finetuning, especially for long-horizon tasks.

Who benefits

AI DevelopmentRoboticsGamingContent Creation

Summary

Researchers introduce SP3O, a novel reward-model-free, critic-free, and gradient-based reinforcement learning algorithm that uses segment-level preference feedback. This method, compatible with stochastic environments and long-horizon tasks, outperforms existing preference-based RL/RLHF algorithms in robotic control and LLM finetuning.

This paper presents SP3O (Segment Pairwise Proximal Policy Optimization), a new algorithm for preference-based reinforcement learning (PbRL) that eliminates the need for a separate reward model. Unlike many existing methods, SP3O is designed for general stochastic environments and can leverage segment-level preferences, which are often easier for human evaluators to provide than full trajectory preferences. SP3O is also critic-free and gradient-based, offering faster convergence than gradient-free alternatives. It constructs an accurate policy value difference estimator using off-policy importance sampling, which then informs a PPO-type loss function for policy gradient computation. The research provides theoretical backing for the algorithm, analyzes segment length tradeoffs, and demonstrates its superior performance in robotic control and large language model fine-tuning, particularly for tasks requiring long-term planning.

Why it matters

SP3O simplifies the process of training AI models from human preferences, making it more efficient and scalable, especially for complex tasks where defining explicit reward functions is difficult or human feedback is best provided on shorter segments.

How to implement this in your domain

  1. 1Experiment with SP3O for fine-tuning large language models where human preference data is available at the segment level.
  2. 2Apply SP3O to robotic control tasks that are difficult to define with traditional reward functions, leveraging human demonstrations or preferences.
  3. 3Evaluate the trade-offs of segment length in preference collection to optimize human annotation effort and model performance.
  4. 4Consider integrating SP3O into existing RLHF pipelines to potentially reduce the complexity and computational cost associated with reward modeling.

Original post by Evan Assmus, Qining Zhang, Lei Ying

"arXiv:2608.02951v1 Announce Type: new Abstract: Preference-based reinforcement learning (PbRL) for general stochastic MDPs often requires training a reward model. Existing reward-model-free methods are either restricted to bandits or deterministic MDPs, such as DPO or P3O, or use…"

View on X

Originally posted by Evan Assmus, Qining Zhang, Lei Ying on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses