English

$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training

Machine Learning 2025-10-21 v2 Artificial Intelligence Computation and Language

Abstract

Reinforcement learning (RL) post-training is crucial for LLM alignment and reasoning, but existing policy-based methods, such as PPO and DPO, can fall short of fixing shortcuts inherited from pre-training. In this work, we introduce QQ\sharp, a value-based algorithm for KL-regularized RL that guides the reference policy using the optimal regularized QQ function. We propose to learn the optimal QQ function using distributional RL on an aggregated online dataset. Unlike prior value-based baselines that guide the model using unregularized QQ-values, our method is theoretically principled and provably learns the optimal policy for the KL-regularized RL problem. Empirically, QQ\sharp outperforms prior baselines in math reasoning benchmarks while maintaining a smaller KL divergence to the reference policy. Theoretically, we establish a reduction from KL-regularized RL to no-regret online learning, providing the first bounds for deterministic MDPs under only realizability. Thanks to distributional RL, our bounds are also variance-dependent and converge faster when the reference policy has small variance. In sum, our results highlight QQ\sharp as an effective approach for post-training LLMs, offering both improved performance and theoretical guarantees. The code can be found at https://github.com/jinpz/q_sharp.

Keywords

Cite

@article{arxiv.2502.20548,
  title  = {$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training},
  author = {Jin Peng Zhou and Kaiwen Wang and Jonathan Chang and Zhaolin Gao and Nathan Kallus and Kilian Q. Weinberger and Kianté Brantley and Wen Sun},
  journal= {arXiv preprint arXiv:2502.20548},
  year   = {2025}
}

Comments

NeurIPS 2025

R2 v1 2026-06-28T22:00:54.623Z