Mind the Gap: Data Rewriting for Stable Off-Policy Supervised Fine-Tuning

• URL: http://arxiv.org/abs/2509.15157v2
• Date: Fri, 19 Sep 2025 03:36:52 GMT
• Title: Mind the Gap: Data Rewriting for Stable Off-Policy Supervised Fine-Tuning
• Authors: Shiwan Zhao, Xuyang Zhao, Jiaming Zhou, Aobo Kong, Qicheng Li, Yong Qin,
• Abstract summary: Supervised fine-tuning (SFT) of large language models can be viewed as an off-policy learning problem.<n>Existing methods mitigate this issue with KL penalties or clipping, which passively updates rather than actively reducing the gap.<n>We propose a simple yet effective data rewriting framework that proactively shrinks the policy gap before training.
• Score: 33.899779762210976
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Supervised fine-tuning (SFT) of large language models can be viewed as an off-policy learning problem, where expert demonstrations come from a fixed behavior policy while training aims to optimize a target policy. Importance sampling is the standard tool for correcting this distribution mismatch, but large policy gaps lead to skewed weights, high variance, and unstable optimization. Existing methods mitigate this issue with KL penalties or clipping, which passively restrict updates rather than actively reducing the gap. We propose a simple yet effective data rewriting framework that proactively shrinks the policy gap before training. For each problem, correct model-generated solutions are kept as on-policy data, while incorrect ones are rewritten through guided re-solving, falling back to expert demonstrations only when needed. This aligns the training distribution with the target policy, reducing variance and improving stability. To handle residual mismatch after rewriting, we additionally apply importance sampling during training, forming a two-stage approach that combines data-level alignment with lightweight optimization-level correction. Experiments on five mathematical reasoning benchmarks show consistent and significant gains over both vanilla SFT and the state-of-the-art Dynamic Fine-Tuning (DFT) approach. Data and code will be released at https://github.com/NKU-HLT/Off-Policy-SFT.

Related papers

• GradAlign: Gradient-Aligned Data Selection for LLM Reinforcement Learning [55.03441672267886]
  We propose GradAlign, a gradient-aligned data selection method for reinforcement learning.<n>We evaluate GradAlign across three data regimes: unreliable reward signals, distribution imbalance, and low-utility training corpus.
  arXiv  Detail & Related papers  (2026-02-25T01:54:50Z)
• Patch the Distribution Mismatch: RL Rewriting Agent for Stable Off-Policy SFT [13.387535599778305]
  We propose a data-centric approach that rewrites downstream training data prior to supervised fine-tuning (SFT)<n>We learn a rewriting policy that better matches the backbone's QA-style generation distribution while preserving diversity.<n>Our method achieves downstream gains comparable to standard SFT while reducing forgetting on non-downstream benchmarks by 12.34% on average.
  arXiv  Detail & Related papers  (2026-02-11T11:51:37Z)
• VESPO: Variational Sequence-Level Soft Policy Optimization for Stable Off-Policy LLM Training [18.849117699859622]
  Training stability is a central challenge in reinforcement learning for large language models.<n>We propose Variational sEquence-level Soft Policy Optimization (VESPO)<n> Experiments on mathematical reasoning benchmarks show that VESPO maintains stable training under staleness ratios up to 64x and fully asynchronous execution.
  arXiv  Detail & Related papers  (2026-02-11T09:48:08Z)
• A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization [58.116300485427764]
  Reinforcement learning post-training can elicit reasoning behaviors in large language models.<n> token-level correction often leads to unstable training dynamics when the degree of off-policyness is large.<n>We propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO)
  arXiv  Detail & Related papers  (2026-01-30T08:47:19Z)
• Ratio-Variance Regularized Policy Optimization for Efficient LLM Fine-tuning [48.34492357368989]
  We propose a primal-dual framework that supports stable on-policy learning and enables principled off-policy data reuse.<n>$R2VPO$ achieves superior performance with average relative gains of up to 17% over strong clipping-based baselines.
  arXiv  Detail & Related papers  (2026-01-06T14:01:42Z)
• Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning [77.92320830700797]
  Reinforcement Learning has played a central role in enabling reasoning capabilities of Large Language Models.<n>We propose a tractable computational framework that tracks and leverages curvature information during policy updates.<n>The algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out.
  arXiv  Detail & Related papers  (2025-10-01T12:29:32Z)
• Relative Entropy Pathwise Policy Optimization [56.86405621176669]
  We show how to construct a value-gradient driven, on-policy algorithm that allow training Q-value models purely from on-policy data.<n>We propose Relative Entropy Pathwise Policy Optimization (REPPO), an efficient on-policy algorithm that combines the sample-efficiency of pathwise policy gradients with the simplicity and minimal memory footprint of standard on-policy learning.
  arXiv  Detail & Related papers  (2025-07-15T06:24:07Z)
• Divergence-Augmented Policy Optimization [12.980566919112034]
  This paper introduces a method to stabilize policy optimization when off-policy data are reused.<n>The idea is to include a Bregman divergence between the behavior policy that generates the data and the current policy to ensure small and safe policy updates with off-policy data.<n> Empirical experiments on Atari games show that in the data-scarce scenario where the reuse of off-policy data becomes necessary, our method can achieve better performance than other state-of-the-art deep reinforcement learning algorithms.
  arXiv  Detail & Related papers  (2025-01-25T02:35:46Z)
• Projected Off-Policy Q-Learning (POP-QL) for Stabilizing Offline Reinforcement Learning [57.83919813698673]
  Projected Off-Policy Q-Learning (POP-QL) is a novel actor-critic algorithm that simultaneously reweights off-policy samples and constrains the policy to prevent divergence and reduce value-approximation error. In our experiments, POP-QL not only shows competitive performance on standard benchmarks, but also out-performs competing methods in tasks where the data-collection policy is significantly sub-optimal.
  arXiv  Detail & Related papers  (2023-11-25T00:30:58Z)
• Off-Policy Evaluation for Large Action Spaces via Policy Convolution [60.6953713877886]
  Policy Convolution family of estimators uses latent structure within actions to strategically convolve the logging and target policies. Experiments on synthetic and benchmark datasets demonstrate remarkable mean squared error (MSE) improvements when using PC.
  arXiv  Detail & Related papers  (2023-10-24T01:00:01Z)
• Offline Imitation Learning with Suboptimal Demonstrations via Relaxed Distribution Matching [109.5084863685397]
  offline imitation learning (IL) promises the ability to learn performant policies from pre-collected demonstrations without interactions with the environment. We present RelaxDICE, which employs an asymmetrically-relaxed f-divergence for explicit support regularization. Our method significantly outperforms the best prior offline method in six standard continuous control environments.
  arXiv  Detail & Related papers  (2023-03-05T03:35:11Z)
• Latent-Variable Advantage-Weighted Policy Optimization for Offline RL [70.01851346635637]
  offline reinforcement learning methods hold the promise of learning policies from pre-collected datasets without the need to query the environment for new transitions. In practice, offline datasets are often heterogeneous, i.e., collected in a variety of scenarios. We propose to leverage latent-variable policies that can represent a broader class of policy distributions. Our method improves the average performance of the next best-performing offline reinforcement learning methods by 49% on heterogeneous datasets.
  arXiv  Detail & Related papers  (2022-03-16T21:17:03Z)
• DDPG++: Striving for Simplicity in Continuous-control Off-Policy Reinforcement Learning [95.60782037764928]
  We show that simple Deterministic Policy Gradient works remarkably well as long as the overestimation bias is controlled. Second, we pinpoint training instabilities, typical of off-policy algorithms, to the greedy policy update step. Third, we show that ideas in the propensity estimation literature can be used to importance-sample transitions from replay buffer and update policy to prevent deterioration of performance.
  arXiv  Detail & Related papers  (2020-06-26T20:21:12Z)

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.