The paper investigates how regularized importance sampling in policy gradient estimation breaks reward invariance and proposes using the covariance between reward and importance-weighted score to restore it. It compares two centering techniques—group-mean baseline and score centering—showing they differ in the mean subtracted and yield distinct unbiased covariance estimators with different variances. The study finds that for sequence-level importance sampling, score centering offers no added benefit beyond the group-mean baseline, whereas for token-level sampling, omitting score centering can introduce bias when the importance ratio is regularized.
arXiv:2607. 04728v1 Announce Type: cross Abstract: Reinforcement learning (RL) post-training for large language models (LLMs) follows a efficient paradigm of "rollout then update", which inevitably results in off-policy training data.
Reinforcement learning (RL) post-training for large language models (LLMs) follows a efficient paradigm of "rollout then update", which inevitably results in off-policy training data. To resolve this, Importance sampling (IS) is proposed, while the token-level ratios compound over long sequences, causing severe variance exploded.
The paper investigates why Group Relative Policy Optimization (GRPO) benefits from per‑prompt normalization by examining the local curvature of the sequence‑level policy gradient. It shows that standard deviation normalization acts as an adaptive gradient, yielding a provably faster convergence rate than unnormalized REINFORCE under mild conditions, with the improvement tied to the average within‑prompt reward standard deviation. The authors also propose IS‑GRPO, an importance‑sampling variant that maintains alignment with the full gradient and offers a tighter convergence guarantee, and empirically validate these theoretical insights on GSM8K and MATH datasets at 1.5B and 7B model scales.
The paper introduces a method called score centering to address the training‑inference mismatch (TIM) that destabilizes reinforcement learning for large language models. By adding an additive correction term that cancels drift between training and inference engines, score centering stabilizes RL and can match or surpass importance‑sampling techniques, especially as model size and mismatch severity increase. The approach also composes with importance sampling, yielding further performance gains in staleness experiments.
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:2609.36816v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-poli...
arXiv:2602. 18037v2 Announce Type: replace-cross Abstract: Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs).
GVPO: Group Variance Policy Optimization is a new post‑training method for large language models that integrates the analytical solution to KL‑constrained reward maximization into its gradient weights. It guarantees a unique optimal solution and allows flexible sampling distributions, avoiding the on‑policy and importance sampling limitations of previous techniques. The approach offers a theoretical guarantee while remaining practically adaptable, positioning it as a reliable and versatile paradigm for LLM post‑training.
The paper introduces POISE, a reinforcement learning algorithm that uses a model’s internal states as a value estimator to reduce variance in reinforcement learning with verifiable rewards (RLVR). By employing a lightweight probe that reads internal signals during the forward pass, POISE predicts baselines online and uses a cross‑rollout construction to keep gradients unbiased. Experiments on Qwen3‑4B and OLMo3‑7B‑Instruct‑DPO across six domains show POISE outperforms existing RLVR baselines, offering more stable training and a value model that generalizes across tasks and scales with the policy.
arXiv:2607. 25091v1 Announce Type: new Abstract: The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated.
Gibbs‑Gap Policy Optimization (GGPO) is a new off‑policy reinforcement learning method that avoids high‑variance learner‑to‑behavior importance weights by minimizing a log‑mean‑exp contrast of KL‑adjusted rewards across responses to the same prompt. Derived from KL‑regularized policy optimization, GGPO guarantees that a learner’s single‑response expected reward is at least the behavior policy’s best‑of‑K reward minus bounded terms, and it includes a reference‑refresh mechanism that empirically improves performance. Experiments on benchmark problems and mathematical reasoning tasks show GGPO outperforming GRPO and AGRO, especially under training‑inference mismatch and with fixed offline data. whyItMatters":"The paper demonstrates a practical improvement in off‑policy RL for language models by reducing variance and ensuring better reward guarantees, which is directly supported by the reported empirical gains and theoretical bounds."
arXiv:2602. 19327v3 Announce Type: replace Abstract: A significant portion of recent research on Large Language Model (LLM) alignment focuses on developing new policy optimization methods based on Group Relative Policy Optimization (GRPO).