Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning

• URL: http://arxiv.org/abs/2504.13818v3
• Date: Wed, 01 Oct 2025 15:58:47 GMT
• Title: Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning
• Authors: Yixuan Even Xu, Yash Savani, Fei Fang, J. Zico Kolter,
• Abstract summary: Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models.<n>It faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive.<n>We introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts.
• Score: 55.15106182268834
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models. However, it faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts, maintaining learning quality while dramatically reducing update costs. We propose a principled subset selection criterion, max-variance down-sampling, that maximizes reward diversity, and provide an efficient $O(n\log n)$ implementation. Empirically, Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least $\mathbf{1.7\times}$ faster across the different reasoning benchmarks and hardware configurations we tested.

Related papers

• Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs [19.079556051442168]
  Reinforcement learning (RL) is widely used to improve large language models on reasoning tasks.<n>But for widely adopted critic-free policy-gradient methods such as REINFORCE and GRPO, high asynchrony makes the policy-gradient estimator markedly noisy.<n>We propose a stabilization method for REINFORCE/ GRPO-style algorithms that scales learning rate based on effective sample size to dampen unreliable updates.
  arXiv  Detail & Related papers  (2026-02-19T18:40:51Z)
• Jackpot: Optimal Budgeted Rejection Sampling for Extreme Actor-Policy Mismatch Reinforcement Learning [31.843129392507716]
  Jackpot is a framework that integrates a unified training objective that jointly updates the policy and rollout models.<n>Our theoretical analysis shows that OBRS consistently moves the rollout distribution closer to the target distribution under a controllable acceptance budget.
  arXiv  Detail & Related papers  (2026-02-05T18:57:01Z)
• Rethinking the Trust Region in LLM Reinforcement Learning [72.25890308541334]
  Proximal Policy Optimization (PPO) serves as the de facto standard algorithm for Large Language Models (LLMs)<n>We propose Divergence Proximal Policy Optimization (DPPO), which substitutes clipping with a more principled constraint.<n>DPPO achieves superior training and efficiency compared to existing methods, offering a more robust foundation for RL-based fine-tuning.
  arXiv  Detail & Related papers  (2026-02-04T18:59:04Z)
• SOUP: Token-level Single-sample Mix-policy Reinforcement Learning for Large Language Models [67.41779761651924]
  SOUP is a framework that unifies off- and on-policy learning within individual samples at the token level.<n>It consistently outperforms standard on-policy training and existing off-policy extensions.
  arXiv  Detail & Related papers  (2026-01-29T09:56:15Z)
• GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization [133.27496265096445]
  We show how to apply Group Relative Policy Optimization under multi-reward setting without examining its suitability.<n>We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues.<n>GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.
  arXiv  Detail & Related papers  (2026-01-08T18:59:24Z)
• Learning More with Less: A Dynamic Dual-Level Down-Sampling Framework for Efficient Policy Optimization [42.2119634259269]
  Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly.<n>We propose the textbfDynamic Dual-Level Down-Sampling (D$3$S) framework that prioritizes the most informative samples and tokens across groups to improve the efficient of policy optimization.
  arXiv  Detail & Related papers  (2025-09-26T09:36:53Z)
• Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions [0.5416466085090772]
  We introduce emphQuantile Reward Policy Optimization (QRPO), which learns from pointwise absolute rewards.<n> QRPO uses quantile rewards to enable regression to the closed-form solution of the KL-regularized RL objective.<n>It consistently achieves top performance on chat and coding evaluations.
  arXiv  Detail & Related papers  (2025-07-10T17:56:24Z)
• Accelerating RL for LLM Reasoning with Optimal Advantage Regression [52.0792918455501]
  We propose a novel two-stage policy optimization framework that directly approximates the optimal advantage function.<n>$A$*-PO achieves competitive performance across a wide range of mathematical reasoning benchmarks.<n>It reduces training time by up to 2$times$ and peak memory usage by over 30% compared to PPO, GRPO, and REBEL.
  arXiv  Detail & Related papers  (2025-05-27T03:58:50Z)
• VerIPO: Cultivating Long Reasoning in Video-LLMs via Verifier-Gudied Iterative Policy Optimization [59.39976343879587]
  VerIPO aims to gradually improve video LLMs' capacity for generating deep, long-term reasoning chains.<n>The training loop benefits from GRPO's expansive search and DPO's targeted optimization.<n>Our trained models exceed the direct inference of large-scale instruction-tuned Video-LLMs.
  arXiv  Detail & Related papers  (2025-05-25T06:41:28Z)
• Kalman Filter Enhanced GRPO for Reinforcement Learning-Based Language Model Reasoning [11.708197376569016]
  Group Relative Policy Optimization ( GRPO) is proposed to compute the advantage for each output by subtracting the mean reward, as the baseline, for all outputs in the group.<n>It can lead to inaccurate advantage estimates in environments with highly noisy rewards, potentially introducing bias.<n>We propose a model, called Kalman Filter Enhanced Group Relative Policy Optimization (KRPO), by using lightweight Kalman filtering to dynamically estimate the latent reward mean and variance.
  arXiv  Detail & Related papers  (2025-05-12T13:09:49Z)
• Do We Truly Need So Many Samples? Multi-LLM Repeated Sampling Efficiently Scales Test-Time Compute [55.330813919992465]
  This paper presents a simple, effective, and cost-efficient strategy to improve LLM performance by scaling test-time compute. Our strategy builds upon the repeated-sampling-then-voting framework, with a novel twist: incorporating multiple models, even weaker ones, to leverage their complementary strengths.
  arXiv  Detail & Related papers  (2025-04-01T13:13:43Z)
• Collab: Controlled Decoding using Mixture of Agents for LLM Alignment [90.6117569025754]
  Reinforcement learning from human feedback has emerged as an effective technique to align Large Language models. Controlled Decoding provides a mechanism for aligning a model at inference time without retraining. We propose a mixture of agent-based decoding strategies leveraging the existing off-the-shelf aligned LLM policies.
  arXiv  Detail & Related papers  (2025-03-27T17:34:25Z)
• RL-finetuning LLMs from on- and off-policy data with a single algorithm [53.70731390624718]
  We introduce a novel reinforcement learning algorithm (AGRO) for fine-tuning large-language models. AGRO leverages the concept of generation consistency, which states that the optimal policy satisfies the notion of consistency across any possible generation of the model. We derive algorithms that find optimal solutions via the sample-based policy gradient and provide theoretical guarantees on their convergence.
  arXiv  Detail & Related papers  (2025-03-25T12:52:38Z)
• Representation-based Reward Modeling for Efficient Safety Alignment of Large Language Model [84.00480999255628]
  Reinforcement Learning algorithms for safety alignment of Large Language Models (LLMs) encounter the challenge of distribution shift. Current approaches typically address this issue through online sampling from the target policy. We propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals.
  arXiv  Detail & Related papers  (2025-03-13T06:40:34Z)
• SDPO: Segment-Level Direct Preference Optimization for Social Agents [56.970902914217156]
  Social agents powered by large language models (LLMs) can simulate human social behaviors but fall short in handling complex social dialogues.<n>We propose Segment-Level Direct Preference Optimization (SDPO), which dynamically select key segments within interactions to optimize multi-turn agent behavior.
  arXiv  Detail & Related papers  (2025-01-03T14:09:46Z)
• A Distribution-Aware Flow-Matching for Generating Unstructured Data for Few-Shot Reinforcement Learning [1.0709300917082865]
  We introduce a distribution-aware flow matching approach to generate synthetic unstructured data for few-shot reinforcement learning. Our approach addresses key challenges in traditional model-based RL, such as overfitting and data correlation. Results demonstrate that our method achieves stable convergence in terms of maximum Q-value while enhancing frame rates by 30% in the initial timestamps.
  arXiv  Detail & Related papers  (2024-09-21T15:50:59Z)
• Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization [75.1240295759264]
  We propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC.<n>We increase the consistency and informativeness of the pairwise preference signals through targeted modifications.<n>We identify that DPO alone is insufficient to model these correlations and capture nuanced variations.
  arXiv  Detail & Related papers  (2024-08-14T11:29:47Z)
• SAPG: Split and Aggregate Policy Gradients [37.433915947580076]
  We propose a new on-policy RL algorithm that can effectively leverage large-scale environments by splitting them into chunks and fusing them back together via importance sampling. Our algorithm, termed SAPG, shows significantly higher performance across a variety of challenging environments where vanilla PPO and other strong baselines fail to achieve high performance.
  arXiv  Detail & Related papers  (2024-07-29T17:59:50Z)
• REBEL: Reinforcement Learning via Regressing Relative Rewards [59.68420022466047]
  We propose REBEL, a minimalist RL algorithm for the era of generative models.<n>In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL.<n>We find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO.
  arXiv  Detail & Related papers  (2024-04-25T17:20:45Z)
• Token-level Direct Preference Optimization [8.249403373337024]
  Fine-tuning pre-trained Large Language Models is essential to align them with human values and intentions. We introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level.
  arXiv  Detail & Related papers  (2024-04-18T08:49:38Z)
• Fine-Tuning Language Models with Reward Learning on Policy [68.70065254564642]
  Reinforcement learning from human feedback (RLHF) has emerged as an effective approach to aligning large language models (LLMs) to human preferences. Despite its popularity, (fixed) reward models may suffer from inaccurate off-distribution. We propose reward learning on policy (RLP), an unsupervised framework that refines a reward model using policy samples to keep it on-distribution.
  arXiv  Detail & Related papers  (2024-03-28T10:02:10Z)
• Active Preference Optimization for Sample Efficient RLHF [27.772423917657626]
  Large Language Models (LLMs) aligned using Reinforcement Learning from Human Feedback (RLHF)<n>We show that uniform sampling of contexts could lead to a policy that suffers a constant sub-optimality gap from the optimal policy.<n>We propose an algorithm, $textttAPO$, that iteratively collects preferences for the most uncertain contexts.
  arXiv  Detail & Related papers  (2024-02-16T08:19:34Z)
• Amortizing intractable inference in large language models [56.92471123778389]
  We use amortized Bayesian inference to sample from intractable posterior distributions. We empirically demonstrate that this distribution-matching paradigm of LLM fine-tuning can serve as an effective alternative to maximum-likelihood training. As an important application, we interpret chain-of-thought reasoning as a latent variable modeling problem.
  arXiv  Detail & Related papers  (2023-10-06T16:36:08Z)
• Optimizing Hyperparameters with Conformal Quantile Regression [7.316604052864345]
  We propose to leverage conformalized quantile regression which makes minimal assumptions about the observation noise. This translates to quicker HPO convergence on empirical benchmarks.
  arXiv  Detail & Related papers  (2023-05-05T15:33:39Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.