Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model

• URL: http://arxiv.org/abs/2507.06892v3
• Date: Fri, 11 Jul 2025 10:32:34 GMT
• Title: Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
• Authors: Jing Liang, Hongyao Tang, Yi Ma, Jinyi Liu, Yan Zheng, Shuyue Hu, Lei Bai, Jianye Hao,
• Abstract summary: We propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix) to enable on-policyRFT methods like PPO and GRPO to leverage off-policy data.<n>ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady improvement.
• Score: 56.92219181993453
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.

Related papers

• Off-Policy Corrected Reward Modeling for Reinforcement Learning from Human Feedback [52.1410307583181]
  We useReinforcement Learning from Human Feedback to train language models (LMs) to follow complex human preferences.<n>As training progresses, the responses generated by the LM no longer resemble the responses seen by the reward model (RM)<n>We propose Off-Policy Corrected Reward Modeling to correct the RM using importance weighting, without requiring new labels or samples.
  arXiv  Detail & Related papers  (2025-07-21T11:19:04Z)
• Ring-lite: Scalable Reasoning via C3PO-Stabilized Reinforcement Learning for LLMs [51.21041884010009]
  Ring-lite is a Mixture-of-Experts (MoE)-based large language model optimized via reinforcement learning (RL)<n>Our approach matches the performance of state-of-the-art (SOTA) small-scale reasoning models on challenging benchmarks.
  arXiv  Detail & Related papers  (2025-06-17T17:12:34Z)
• Mutual-Taught for Co-adapting Policy and Reward Models [43.11214888109746]
  We propose Mutual-Taught, a self-training method that iteratively improves both the policy model and the reward model.<n> Experimental results demonstrate that this iterative approach leads to consistent improvements in both models.
  arXiv  Detail & Related papers  (2025-05-17T04:34:23Z)
• On the Robustness of Reward Models for Language Model Alignment [9.804782604188656]
  We study the cause of over-optimization in reward models trained with the Bradley-Terry (BT) model.<n>We show that the excessive dispersion of hidden state norms is the main source of over-optimization.<n>We apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale.
  arXiv  Detail & Related papers  (2025-05-12T06:48:26Z)
• Lean and Mean: Decoupled Value Policy Optimization with Global Value Guidance [52.65461207786633]
  Policy-based Reinforcement Learning from Human Feedback is essential for aligning large language models with human preferences.<n>It requires joint training of an actor and critic with a pretrained, fixed reward model for guidance.<n>We propose textbfDecoupled Value Policy Optimization (DVPO), a lean framework that replaces traditional reward modeling with a pretrained emphglobal value model (GVM)
  arXiv  Detail & Related papers  (2025-02-24T08:11:33Z)
• Kimi k1.5: Scaling Reinforcement Learning with LLMs [84.95584393629998]
  We report on the training practice of Kimi k1.5, our latest multi-modal language model trained with reinforcement learning.<n>Long context scaling and improved policy optimization methods are key ingredients of our approach.<n>Our system achieves state-of-the-art reasoning performance across multiple benchmarks and modalities.
  arXiv  Detail & Related papers  (2025-01-22T02:48:14Z)
• RRM: Robust Reward Model Training Mitigates Reward Hacking [51.12341734942797]
  Reward models (RMs) play a pivotal role in aligning large language models with human preferences.<n>We introduce a causal framework that learns preferences independent of these artifacts.<n>Experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model.
  arXiv  Detail & Related papers  (2024-09-20T01:46:07Z)
• Self-Play Preference Optimization for Language Model Alignment [75.83359213697854]
  Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences. We propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game. Our approach, dubbed Self-Play Preference Optimization (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium.
  arXiv  Detail & Related papers  (2024-05-01T17:59:20Z)
• ReMax: A Simple, Effective, and Efficient Reinforcement Learning Method for Aligning Large Language Models [30.276168676690045]
  Reinforcement Learning from Human Feedback (RLHF) is key to aligning Large Language Models (LLMs) We present ReMax, which leverages 3 properties of RLHF: fast simulation, deterministic transitions, and trajectory-level rewards. It is simpler to implement, eliminates more than 4 hyper- parameters in PPO, reduces GPU memory usage, and shortens training time. Applying ReMax to a Mistral-7B model resulted in a 94.78% win rate on the AlpacaEval leaderboard and a 7.739 score on MT-bench, setting a new SOTA for open-source 7B models
  arXiv  Detail & Related papers  (2023-10-16T15:25:14Z)

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.