The paper proposes a new technique that averages the logits of a frozen reference policy (such as one obtained via supervised fine‑tuning) and a trainable policy, integrating this approach into Group Relative Policy Optimization (GRPO). Unlike methods that use Kullback–Leibler regularization or a critic, the proposed method couples the policies solely through logit averaging, aiming to combine the reasoning strengths of the trainable policy with the formatting benefits of the reference policy. Experiments on MATH, cn‑k12, and MMLU demonstrate that this approach achieves higher or comparable accuracy to the standard KL‑regularized GRPO.
arXiv:2603. 00963v2 Announce Type: replace Abstract: While reinforcement learning (RL) has been central to the recent success of large language models (LLMs), RL optimization is notoriously unstable, especially when compared to supervised fine-tuning (SFT).
arXiv:2606. 20008v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards has become a central tool for improving the reasoning ability of large language models, but current methods face a trade-off between simplicity and credit assignment.
The paper introduces Best‑Practice Critic Optimization (BPCO), a stable and efficient recipe for training a critic in reinforcement learning for large language models. BPCO combines DPPO, bounded value predictions, Monte Carlo targets, unnormalized policy advantages, and length‑adaptive advantage estimation, allowing the critic to be conditioned on hidden reward information. Experiments on mathematical reasoning tasks with models from 1.5B to 30B parameters show that BPCO consistently outperforms a strong critic‑based baseline and matches or exceeds group‑based methods while sampling only one response per prompt.
arXiv:2608.23566v2 Announce Type: replace-cross Abstract: Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for...
arXiv:2607. 18470v1 Announce Type: cross Abstract: Group Relative Policy Optimization (GRPO) has shown strong effectiveness in reinforcement learning from verifiable feedback, where sampled rollouts can be compared within a group using task-provided correctness signals.
arXiv:2608. 16739v1 Announce Type: new Abstract: Reinforcement learning algorithms for Large Language Models (LLMs) are largely distinguished by their variance reduction strategy.
arXiv:2609.26355v1 Announce Type: new Abstract: Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted ma...
arXiv:2606. 28707v1 Announce Type: new Abstract: Critic-free reinforcement learning with verifiable rewards (RLVR), exemplified by Group Relative Policy Optimization (GRPO), avoids training a value function (critic) and reduces memory and compute overhead relative to critic-based PPO pipelines for aligning large language models.
arXiv:2606. 16154v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but GRPO-style optimization remains prone to collapse.
arXiv:2604. 13780v2 Announce Type: replace-cross Abstract: Soft Q-learning has emerged as a versatile model-free method for entropy-regularised reinforcement learning, optimising for returns augmented with a penalty on the divergence from a reference policy.
arXiv:2606. 25451v1 Announce Type: new Abstract: Estimating token-level advantages in reinforcement learning (RL) for language models remains challenging because scaling up episodic experience collection is expensive.
arXiv:2606. 08346v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving the reasoning capabilities of large language models (LLMs).