Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning

• URL: http://arxiv.org/abs/2604.08926v1
• Date: Fri, 10 Apr 2026 03:42:16 GMT
• Title: Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning
• Abstract summary: Post-training paradigms for Large Language Models (LLMs) face a fundamental dilemma.<n>SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gradient variance.<n>Existing unified optimization strategies often employ naive loss weighting, overlooking the statistical conflict between these distinct gradient signals.<n>We propose textbfDYPO (Dynamic Policy Optimization), a unified framework designed to structurally mitigate this conflict.
• Score: 10.4404175363648
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Post-training paradigms for Large Language Models (LLMs), primarily Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), face a fundamental dilemma: SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gradient variance. Existing unified optimization strategies often employ naive loss weighting, overlooking the statistical conflict between these distinct gradient signals. In this paper, we provide a rigorous theoretical analysis of this bias-variance trade-off and propose \textbf{DYPO} (Dynamic Policy Optimization), a unified framework designed to structurally mitigate this conflict. DYPO integrates three core components: (1) a \textit{Group Alignment Loss (GAL)} that leverages intrinsic group dynamics to significantly reduce RL gradient variance; (2) a \textit{Multi-Teacher Distillation} mechanism that corrects SFT fitting bias via diverse reasoning paths; and (3) a \textit{Dynamic Exploitation-Exploration Gating} mechanism that adaptively arbitrates between stable SFT and exploratory RL based on reward feedback. Theoretical analysis confirms that DYPO linearly reduces fitting bias and minimizes overall variance. Extensive experiments demonstrate that DYPO significantly outperforms traditional sequential pipelines, achieving an average improvement of 4.8\% on complex reasoning benchmarks and 13.3\% on out-of-distribution tasks. Our code is publicly available at https://github.com/Tocci-Zhu/DYPO.

Related papers

• SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs [55.96262261252617]
  Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs)<n>Our preliminary experiments revealed a phenomenon: SFT suffers from severe task conflicts under multi-stage training, whereas RL enables stable coexistence across diverse tasks.<n>We propose Parallel-RL, a paradigm that decouples multi-task training, significantly improving efficiency and flexibility.
  arXiv  Detail & Related papers  (2026-08-04T12:32:26Z)
• DISA: Offline Importance Sampling for Distribution-Matching LLM-RL [56.9445657766829]
  We introduce DISA, which moves this calibration problem outside the RL loop.<n> DISA draws proposal trajectories offline, estimates the partition function via importance sampling, and freezes the resulting partition-function estimate.<n>On two open-weight backbones across six math and three code benchmarks, DISA matches or exceeds the online-coupled distribution-matching baseline Flow.
  arXiv  Detail & Related papers  (2026-05-17T07:14:44Z)
• Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data [55.84428098924793]
  We propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration.<n>We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance.<n> Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO.
  arXiv  Detail & Related papers  (2026-04-20T16:43:28Z)
• S-GRPO: Unified Post-Training for Large Vision-Language Models [11.901218794806796]
  We propose Supervised Group Relative Policy Optimization (S-GRPO) for adapting Large Vision-Language Models (LVLMs)<n>S-GRPO integrates the guidance of imitation learning into the multi-trajectory exploration of preference optimization.<n>It gracefully bridges the gap between Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL)
  arXiv  Detail & Related papers  (2026-04-17T08:39:07Z)
• GFT: From Imitation to Reward Fine-Tuning with Unbiased Group Advantages and Dynamic Coefficient Rectification [37.24414986596345]
  Group Fine-Tuning (GFT) is a unified post-training framework for large language models.<n>GFT consistently surpasses SFT-based methods and yields policies that integrate more smoothly with subsequent RL training.
  arXiv  Detail & Related papers  (2026-04-15T15:49:58Z)
• SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning [54.393763477932474]
  Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has emerged as the standard post-training paradigm for large language models (LLMs)<n>We propose SED-SFT, which adaptively encourages diversity based on the token exploration space.<n>This framework introduces a selective entropy regularization term with a selective masking mechanism into the optimization objective.
  arXiv  Detail & Related papers  (2026-02-07T09:39:21Z)
• ConsistentRFT: Reducing Visual Hallucinations in Flow-based Reinforcement Fine-Tuning [85.20505958752928]
  Reinforcement Fine-Tuning (RFT) on flow-based models is crucial for preference alignment.<n>RFT often introduce visual hallucinations like over-optimized details and semantic misalignment.<n>This work preliminarily explores why visual hallucinations arise and how to reduce them.
  arXiv  Detail & Related papers  (2026-02-03T11:49:46Z)
• Trust-Region Adaptive Policy Optimization [82.09255251747818]
  Post-training methods play an important role in improving large language models' (LLMs) complex reasoning abilities.<n>We introduce TRAPO, a framework that interleavesSupervised Fine-Tuning (SFT) and Reinforcement Learning (RL) within each training instance.<n>Experiments on five mathematical reasoning benchmarks show that TRAPO consistently surpasses standard SFT, RL, and SFT-then-RL pipelines.
  arXiv  Detail & Related papers  (2025-12-19T14:37:07Z)
• DaGRPO: Rectifying Gradient Conflict in Reasoning via Distinctiveness-Aware Group Relative Policy Optimization [20.66452395111739]
  We propose Distinctiveness-aware Group Relative Policy Optimization (DaGRPO)<n>DaGRPO incorporates two core mechanisms: (1) Sequence-level Gradient Rectification, which utilizes fine-grained scoring to dynamically mask sample pairs with low distinctiveness; and (2) Off-policy Data Augmentation, which introduces high-quality anchors to recover training signals for challenging tasks.<n>In-depth analysis confirms that DaGRPO effectively mitigates gradient explosion and accelerates the emergence of long-chain reasoning capabilities.
  arXiv  Detail & Related papers  (2025-12-06T07:51:36Z)
• MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources [113.33902847941941]
  Variance-Aware Sampling (VAS) is a data selection strategy guided by Variance Promotion Score (VPS)<n>We release large-scale, carefully curated resources containing 1.6M long CoT cold-start data and 15k RL QA pairs.<n> Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS.
  arXiv  Detail & Related papers  (2025-09-25T14:58:29Z)
• AMFT: Aligning LLM Reasoners by Meta-Learning the Optimal Imitation-Exploration Balance [7.685078284407324]
  Large Language Models (LLMs) are typically fine-tuned for reasoning tasks through a two-stage pipeline Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL)<n>Recent single-stage methods attempt to unify SFT and RL using principleds, but lack a mechanism for dynamically balancing the two paradigms.<n>We introduce textbf Meta Fine-Tuning (AMFT), a novel single-stage algorithm that learns the optimal balance between SFT's implicit, path-level reward and RL's explicit, outcome-based reward.
  arXiv  Detail & Related papers  (2025-08-09T11:40:54Z)
• Implicit Reward as the Bridge: A Unified View of SFT and DPO Connections [65.36449542323277]
  We present a unified theoretical framework bridgingSupervised Fine-Tuning (SFT) and preference learning in Large Language Model (LLM) post-training.<n>We propose a simple yet effective learning rate reduction approach that yields significant performance improvements.
  arXiv  Detail & Related papers  (2025-06-15T05:42:29Z)

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.