Preference Distillation via Value based Reinforcement Learning

• URL: http://arxiv.org/abs/2509.16965v1
• Date: Sun, 21 Sep 2025 07:52:28 GMT
• Title: Preference Distillation via Value based Reinforcement Learning
• Authors: Minchan Kwon, Junwon Ko, Kangil Kim, Junmo Kim,
• Abstract summary: We propose textitTeacher Value-based Knowledge Distillation (TVKD), which introduces an auxiliary reward from the value function of the teacher model to provide a soft guide.<n>TVKD can be integrated into the standard DPO training framework and does not require additional rollouts.<n>Our experimental results show that TVKD consistently improves performance across various benchmarks and model sizes.
• Score: 16.165599808093408
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Direct Preference Optimization (DPO) is a powerful paradigm to align language models with human preferences using pairwise comparisons. However, its binary win-or-loss supervision often proves insufficient for training small models with limited capacity. Prior works attempt to distill information from large teacher models using behavior cloning or KL divergence. These methods often focus on mimicking current behavior and overlook distilling reward modeling. To address this issue, we propose \textit{Teacher Value-based Knowledge Distillation} (TVKD), which introduces an auxiliary reward from the value function of the teacher model to provide a soft guide. This auxiliary reward is formulated to satisfy potential-based reward shaping, ensuring that the global reward structure and optimal policy of DPO are preserved. TVKD can be integrated into the standard DPO training framework and does not require additional rollouts. Our experimental results show that TVKD consistently improves performance across various benchmarks and model sizes.

Related papers

• Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation [57.524909883706556]
  On-policy distillation (OPD) has demonstrated strong empirical gains in improving student performance.<n>This work introduces a flexible reference model and a reward scaling factor that controls the relative weight of the reward term against the KL regularization.<n>In particular, in the setting where we merge the knowledge from different domain experts, obtained by applying domain-specific RL to the same student model, ExOPD enables the student to even surpass the teacher's performance boundary.
  arXiv  Detail & Related papers  (2026-02-12T16:14:29Z)
• GFRIEND: Generative Few-shot Reward Inference through EfficieNt DPO [3.189559302776161]
  The ability to train high-performing reward models with few-shot data is critical for enhancing the efficiency and scalability of Reinforcement Learning from Human Feedback.<n>We propose a data augmentation and expansion framework that enables generative reward models trained on small datasets to achieve comparable performance to those trained on large-scale datasets.
  arXiv  Detail & Related papers  (2025-06-10T16:37:13Z)
• daDPO: Distribution-Aware DPO for Distilling Conversational Abilities [48.745922491268004]
  This paper introduces daDPO (Distribution-Aware DPO), a unified method for preference optimization and distribution-based distillation.<n>We show that daDPO outperforms existing methods in restoring performance for pruned models and enhancing smaller LLM models.
  arXiv  Detail & Related papers  (2025-06-03T03:39:29Z)
• VARD: Efficient and Dense Fine-Tuning for Diffusion Models with Value-based RL [28.95582264086289]
  VAlue-based Reinforced Diffusion (VARD) is a novel approach that first learns a value function predicting expection of rewards from intermediate states.<n>Our method maintains proximity to the pretrained model while enabling effective and stable training via backpropagation.
  arXiv  Detail & Related papers  (2025-05-21T17:44:37Z)
• Active Data Curation Effectively Distills Large-Scale Multimodal Models [66.23057263509027]
  Knowledge distillation (KD) is the de facto standard for compressing large-scale models into smaller ones.<n>In this work we explore an alternative, yet simple approach -- active data curation as effective distillation for contrastive multimodal pretraining.<n>Our simple online batch selection method, ACID, outperforms strong KD baselines across various model-, data- and compute-configurations.
  arXiv  Detail & Related papers  (2024-11-27T18:50:15Z)
• Exploring and Enhancing the Transfer of Distribution in Knowledge Distillation for Autoregressive Language Models [62.5501109475725]
  Knowledge distillation (KD) is a technique that compresses large teacher models by training smaller student models to mimic them. This paper introduces Online Knowledge Distillation (OKD), where the teacher network integrates small online modules to concurrently train with the student model. OKD achieves or exceeds the performance of leading methods in various model architectures and sizes, reducing training time by up to fourfold.
  arXiv  Detail & Related papers  (2024-09-19T07:05:26Z)
• Robust Preference Optimization through Reward Model Distillation [68.65844394615702]
  Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data.<n>We analyze this phenomenon and use distillation to get a better proxy for the true preference distribution over generation pairs.<n>Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations.
  arXiv  Detail & Related papers  (2024-05-29T17:39:48Z)
• RewardBench: Evaluating Reward Models for Language Modeling [100.28366840977966]
  We present RewardBench, a benchmark dataset and code-base for evaluation of reward models. The dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods.
  arXiv  Detail & Related papers  (2024-03-20T17:49:54Z)
• Distilling Privileged Multimodal Information for Expression Recognition using Optimal Transport [46.91791643660991]
  Deep learning models for multimodal expression recognition have reached remarkable performance in controlled laboratory environments. These models struggle in the wild because of the unavailability and quality of modalities used for training. In practice, only a subset of the training-time modalities may be available at test time. Learning with privileged information enables models to exploit data from additional modalities that are only available during training.
  arXiv  Detail & Related papers  (2024-01-27T19:44:15Z)
• Self-Feature Regularization: Self-Feature Distillation Without Teacher Models [0.0]
  Self-Feature Regularization(SFR) is proposed, which uses features in the deep layers to supervise feature learning in the shallow layers. We firstly use generalization-l2 loss to match local features and a many-to-one approach to distill more intensively in the channel dimension.
  arXiv  Detail & Related papers  (2021-03-12T15:29:00Z)
• MixKD: Towards Efficient Distillation of Large-scale Language Models [129.73786264834894]
  We propose MixKD, a data-agnostic distillation framework, to endow the resulting model with stronger generalization ability. We prove from a theoretical perspective that under reasonable conditions MixKD gives rise to a smaller gap between the error and the empirical error. Experiments under a limited-data setting and ablation studies further demonstrate the advantages of the proposed approach.
  arXiv  Detail & Related papers  (2020-11-01T18:47:51Z)

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.