On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting

• URL: http://arxiv.org/abs/2508.11408v2
• Date: Fri, 10 Oct 2025 06:33:12 GMT
• Title: On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting
• Authors: Wenhao Zhang, Yuexiang Xie, Yuchang Sun, Yanxi Chen, Guoyin Wang, Yaliang Li, Bolin Ding, Jingren Zhou,
• Abstract summary: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are two prominent post-training paradigms for refining the capabilities and aligning the behavior of Large Language Models (LLMs)<n>Existing approaches that integrate SFT and RL often face the risk of disrupting established response patterns and inducing overfitting to expert data.<n>We propose CHORD, a framework for Controllable Harmonization of On- and Off-Policy Reinforcement Learning via Dynamic Weighting.
• Score: 91.38734024438357
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are two prominent post-training paradigms for refining the capabilities and aligning the behavior of Large Language Models (LLMs). Existing approaches that integrate SFT and RL often face the risk of disrupting established response patterns and inducing overfitting to expert data. To address this, we present a novel investigation into the unified view of SFT and RL through an off-policy versus on-policy lens. We propose CHORD, a framework for Controllable Harmonization of On- and Off-Policy Reinforcement Learning via Dynamic Weighting, which reframes SFT not as a separate stage but as a dynamically weighted auxiliary objective within the on-policy RL process. Based on an analysis of off-policy expert data's influence at both holistic and granular levels, we incorporate a dual-control mechanism in CHORD. Specifically, the framework first employs a global coefficient to holistically guide the transition from off-policy imitation to on-policy exploration, and then applies a token-wise weighting function that enables granular learning from the expert, which promotes on-policy exploration and mitigates disruption from off-policy data. We conduct extensive experiments on mathematical reasoning problems and practical tool-use tasks, providing empirical evidence that CHORD achieves a stable and efficient learning process. By effectively harmonizing off-policy expert data with on-policy exploration, CHORD demonstrates significant improvements over baselines. We release the implementation at https://github.com/modelscope/Trinity-RFT/tree/main/examples/mix_chord to inspire further research.

Related papers

• Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration [56.074760766965085]
  PRISM achieves a dynamics-aware framework that arbitrates data based on its degree of cognitive conflict with the model's existing knowledge.<n>Our findings suggest that disentangling data based on internal optimization regimes is crucial for scalable and robust agent alignment.
  arXiv  Detail & Related papers  (2026-01-12T05:43:20Z)
• Improving Deepfake Detection with Reinforcement Learning-Based Adaptive Data Augmentation [60.04281435591454]
  CRDA (Curriculum Reinforcement-Learning Data Augmentation) is a novel framework guiding detectors to progressively master multi-domain forgery features.<n>Central to our approach is integrating reinforcement learning and causal inference.<n>Our method significantly improves detector generalizability, outperforming SOTA methods across multiple cross-domain datasets.
  arXiv  Detail & Related papers  (2025-11-10T12:45:52Z)
• Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting [40.80967570661867]
  Adapting language models to new tasks via post-training carries the risk of degrading existing capabilities.<n>We compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL)<n>RL leads to less forgetting than SFT while achieving comparable or higher target task performance.
  arXiv  Detail & Related papers  (2025-10-21T17:59:41Z)
• Group-Relative REINFORCE Is Secretly an Off-Policy Algorithm: Demystifying Some Myths About GRPO and Its Friends [64.71326476563213]
  Off-policy reinforcement learning for large language models (LLMs) is attracting growing interest.<n>We present a first-principles derivation for grouprelative REINFORCE without assuming a specific training data distribution.<n>This perspective yields two general principles for adapting REINFORCE to off-policy settings.
  arXiv  Detail & Related papers  (2025-09-29T02:34:54Z)
• Diffusion Guidance Is a Controllable Policy Improvement Operator [98.11511661904618]
  CFGRL is trained with the simplicity of supervised learning, yet can further improve on the policies in the data.<n>On offline RL tasks, we observe a reliable trend -- increased guidance weighting leads to increased performance.
  arXiv  Detail & Related papers  (2025-05-29T14:06:50Z)
• Preference-Based Multi-Agent Reinforcement Learning: Data Coverage and Algorithmic Techniques [65.55451717632317]
  We study Preference-Based Multi-Agent Reinforcement Learning (PbMARL)<n>We identify the Nash equilibrium from a preference-only offline dataset in general-sum games.<n>Our findings underscore the multifaceted approach required for PbMARL.
  arXiv  Detail & Related papers  (2024-09-01T13:14:41Z)
• SAMBO-RL: Shifts-aware Model-based Offline Reinforcement Learning [9.88109749688605]
  Model-based offline reinforcement learning trains policies using pre-collected datasets and learned environment models.<n>This paper offers a comprehensive analysis that disentangles the problem into two fundamental components: model bias and policy shift.<n>We introduce Shifts-aware Model-based Offline Reinforcement Learning (SAMBO-RL), a practical framework that efficiently trains classifiers to approximate SAR for policy optimization.
  arXiv  Detail & Related papers  (2024-08-23T04:25:09Z)
• Pessimistic Causal Reinforcement Learning with Mediators for Confounded Offline Data [17.991833729722288]
  We propose a novel policy learning algorithm, PESsimistic CAusal Learning (PESCAL) Our key observation is that, by incorporating auxiliary variables that mediate the effect of actions on system dynamics, it is sufficient to learn a lower bound of the mediator distribution function, instead of the Q-function. We provide theoretical guarantees for the algorithms we propose, and demonstrate their efficacy through simulations, as well as real-world experiments utilizing offline datasets from a leading ride-hailing platform.
  arXiv  Detail & Related papers  (2024-03-18T14:51:19Z)
• Jump-Start Reinforcement Learning [68.82380421479675]
  We present a meta algorithm that can use offline data, demonstrations, or a pre-existing policy to initialize an RL policy. In particular, we propose Jump-Start Reinforcement Learning (JSRL), an algorithm that employs two policies to solve tasks. We show via experiments that JSRL is able to significantly outperform existing imitation and reinforcement learning algorithms.
  arXiv  Detail & Related papers  (2022-04-05T17:25:22Z)
• Federated Ensemble Model-based Reinforcement Learning in Edge Computing [21.840086997141498]
  Federated learning (FL) is a privacy-preserving distributed machine learning paradigm. We propose a novel FRL algorithm that effectively incorporates model-based RL and ensemble knowledge distillation into FL for the first time. Specifically, we utilise FL and knowledge distillation to create an ensemble of dynamics models for clients, and then train the policy by solely using the ensemble model without interacting with the environment.
  arXiv  Detail & Related papers  (2021-09-12T16:19:10Z)
• Deep RL With Information Constrained Policies: Generalization in Continuous Control [21.46148507577606]
  We show that a natural constraint on information flow might confer onto artificial agents in continuous control tasks. We implement a novel Capacity-Limited Actor-Critic (CLAC) algorithm. Our experiments show that compared to alternative approaches, CLAC offers improvements in generalization between training and modified test environments.
  arXiv  Detail & Related papers  (2020-10-09T15:42:21Z)

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.