Differential Information Distribution: A Bayesian Perspective on Direct Preference Optimization

• URL: http://arxiv.org/abs/2505.23761v2
• Date: Thu, 02 Oct 2025 17:59:55 GMT
• Title: Differential Information Distribution: A Bayesian Perspective on Direct Preference Optimization
• Authors: Yunjae Won, Hyunji Lee, Hyeonbin Hwang, Minjoon Seo,
• Abstract summary: We study the goal of preference optimization as learning the differential information required to update a reference policy into a target policy.<n>First, we find that DPO's log-ratio reward is uniquely justified when preferences encode the Differential Information needed to update a reference policy into the target policy.<n>Second, we discuss how commonly observed training dynamics in DPO, including changes in log-likelihood and policy exploration, stem from a power-law DID relationship.
• Score: 35.335072390336855
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Direct Preference Optimization (DPO) has been widely used for aligning language models with human preferences in a supervised manner. However, several key questions remain unresolved: the rationale behind its log-ratio reward, how the statistical structure of preference datasets shapes its training dynamics, and how those dynamics impact downstream capabilities. We approach these questions from a Bayesian perspective, interpreting the goal of preference optimization as learning the differential information required to update a reference policy into a target policy. To formalize this view, we introduce the Differential Information Distribution (DID), defined as the distribution over samples that carry the Bayesian evidence required to update policies. We introduce three complementary insights by viewing preference optimization through the DID. First, we find that DPO's log-ratio reward is uniquely justified when preferences encode the Differential Information needed to update a reference policy into the target policy. Second, we discuss how commonly observed training dynamics in DPO, including changes in log-likelihood and policy exploration, stem from a power-law DID relationship. Finally, we analyze how training dynamics influence downstream performance using the entropy of DID, a principled measure of uncertainty in the learned information. We observe that learning high-entropy DID improves open-ended instruction-following, while low-entropy DID benefits knowledge-intensive QA. Taken together, our results show that DPO's reward design, training dynamics, and downstream capabilities all emerge as natural consequences of learning Differential Information, offering both a principled theoretical foundation and practical guidance for preference-based alignment.

Related papers

• 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)
• SGPO: Self-Generated Preference Optimization based on Self-Improver [6.528083376369728]
  Large language models (LLMs) require alignment to human preferences for practical and reliable deployment.<n>We propose Self-Generated Preference Optimization based on Self-Improver (SGPO)<n>The improver refines responses from a policy model to self-generate preference data for direct preference optimization (DPO) of the policy model.<n> Experimental results on AlpacaEval 2.0 and Arena-Hard show that the proposed SGPO significantly improves performance over DPO and baseline self-improving methods.
  arXiv  Detail & Related papers  (2025-07-27T08:55:40Z)
• Understanding the Impact of Sampling Quality in Direct Preference Optimization [4.122673728216191]
  We study how data of higher quality can be leveraged to improve performance in Direct Preference Optimization (DPO)<n>Our analyses show that both the solution space and the convergence behavior of DPO depend on the support and quality of the data-generating distribution.
  arXiv  Detail & Related papers  (2025-06-03T18:12:40Z)
• Leveraging Robust Optimization for LLM Alignment under Distribution Shifts [52.983390470606146]
  Preference alignment methods are increasingly critical for steering large language models to generate outputs consistent with human values.<n>We propose a novel distribution-aware optimization framework that improves preference alignment despite such shifts.
  arXiv  Detail & Related papers  (2025-04-08T09:14:38Z)
• Efficient Safety Alignment of Large Language Models via Preference Re-ranking and Representation-based Reward Modeling [84.00480999255628]
  Reinforcement Learning algorithms for safety alignment of Large Language Models (LLMs) encounter the challenge of distribution shift.<n>Current approaches typically address this issue through online sampling from the target policy.<n>We propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals.
  arXiv  Detail & Related papers  (2025-03-13T06:40:34Z)
• A Survey of Direct Preference Optimization [103.59317151002693]
  Large Language Models (LLMs) have demonstrated unprecedented generative capabilities.<n>Their alignment with human values remains critical for ensuring helpful and harmless deployments.<n>Direct Preference Optimization (DPO) has recently gained prominence as a streamlined alternative.
  arXiv  Detail & Related papers  (2025-03-12T08:45:15Z)
• Active Learning for Direct Preference Optimization [59.84525302418018]
  Direct preference optimization (DPO) is a form of reinforcement learning from human feedback.<n>We propose an active learning framework for DPO, which can be applied to collect human feedback online or to choose the most informative subset of already collected feedback offline.
  arXiv  Detail & Related papers  (2025-03-03T00:36:31Z)
• What Has Been Overlooked in Contrastive Source-Free Domain Adaptation: Leveraging Source-Informed Latent Augmentation within Neighborhood Context [28.634315143647385]
  Source-free domain adaptation (SFDA) involves adapting a model originally trained using a labeled dataset to perform effectively on an unlabeled dataset.<n>This adaptation is especially crucial when significant disparities in data distributions exist between the two domains.<n>We introduce a straightforward yet highly effective latent augmentation method tailored for contrastive SFDA.
  arXiv  Detail & Related papers  (2024-12-18T20:09:46Z)
• BPO: Towards Balanced Preference Optimization between Knowledge Breadth and Depth in Alignment [32.095601071459136]
  We introduce the concepts of knowledge breadth and knowledge depth, which measure the comprehensiveness and depth of a knowledge source.<n>We propose Balanced Preference Optimization (BPO), designed to dynamically augment the knowledge depth of each sample.<n>BPO is motivated by the observation that the usefulness of knowledge varies across samples, necessitating tailored learning of knowledge depth.
  arXiv  Detail & Related papers  (2024-11-16T23:53:27Z)
• Optimal Classification under Performative Distribution Shift [13.508249764979075]
  We propose a novel view in which performative effects are modelled as push-forward measures. We prove the convexity of the performative risk under a new set of assumptions. We also establish a connection with adversarially robust classification by reformulating the minimization of the performative risk as a min-max variational problem.
  arXiv  Detail & Related papers  (2024-11-04T12:20:13Z)
• Forward KL Regularized Preference Optimization for Aligning Diffusion Policies [8.958830452149789]
  A central problem for learning diffusion policies is to align the policy output with human intents in various tasks.<n>We propose a novel framework, Forward KL regularized Preference optimization, to align the diffusion policy with preferences directly.<n>The results show our method exhibits superior alignment with preferences and outperforms previous state-of-the-art algorithms.
  arXiv  Detail & Related papers  (2024-09-09T13:56:03Z)
• Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization [75.1240295759264]
  We propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC.<n>We increase the consistency and informativeness of the pairwise preference signals through targeted modifications.<n>We identify that DPO alone is insufficient to model these correlations and capture nuanced variations.
  arXiv  Detail & Related papers  (2024-08-14T11:29:47Z)
• Stratified Prediction-Powered Inference for Hybrid Language Model Evaluation [62.2436697657307]
  Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data.<n>We propose a method called Stratified Prediction-Powered Inference (StratPPI)<n>We show that the basic PPI estimates can be considerably improved by employing simple data stratification strategies.
  arXiv  Detail & Related papers  (2024-06-06T17:37:39Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences [24.645259298082436]
  We take a step towards a deeper understanding of learning from human preferences by systematically comparing the paradigm of reinforcement learning from human feedback (RLHF) with the recently proposed paradigm of direct preference optimization (DPO) We derive minimax statistical bounds on the suboptimality gap induced by both RLHF and DPO. We extend our analysis to the approximate optimization setting and derive exponentially decaying convergence rates for both RLHF and DPO.
  arXiv  Detail & Related papers  (2024-03-04T09:13:14Z)
• Towards Efficient Exact Optimization of Language Model Alignment [93.39181634597877]
  Direct preference optimization (DPO) was proposed to directly optimize the policy from preference data. We show that DPO derived based on the optimal solution of problem leads to a compromised mean-seeking approximation of the optimal solution in practice. We propose efficient exact optimization (EXO) of the alignment objective.
  arXiv  Detail & Related papers  (2024-02-01T18:51:54Z)
• Statistically Efficient Variance Reduction with Double Policy Estimation for Off-Policy Evaluation in Sequence-Modeled Reinforcement Learning [53.97273491846883]
  We propose DPE: an RL algorithm that blends offline sequence modeling and offline reinforcement learning with Double Policy Estimation. We validate our method in multiple tasks of OpenAI Gym with D4RL benchmarks.
  arXiv  Detail & Related papers  (2023-08-28T20:46:07Z)
• STEERING: Stein Information Directed Exploration for Model-Based Reinforcement Learning [111.75423966239092]
  We propose an exploration incentive in terms of the integral probability metric (IPM) between a current estimate of the transition model and the unknown optimal. Based on KSD, we develop a novel algorithm algo: textbfSTEin information dirtextbfEcted exploration for model-based textbfReinforcement LearntextbfING.
  arXiv  Detail & Related papers  (2023-01-28T00:49:28Z)
• Policy learning "without" overlap: Pessimism and generalized empirical Bernstein's inequality [94.89246810243053]
  This paper studies offline policy learning, which aims at utilizing observations collected a priori to learn an optimal individualized decision rule.<n>Existing policy learning methods rely on a uniform overlap assumption, i.e., the propensities of exploring all actions for all individual characteristics must be lower bounded.<n>We propose Pessimistic Policy Learning (PPL), a new algorithm that optimize lower confidence bounds (LCBs) instead of point estimates.
  arXiv  Detail & Related papers  (2022-12-19T22:43:08Z)
• Unified Instance and Knowledge Alignment Pretraining for Aspect-based Sentiment Analysis [96.53859361560505]
  Aspect-based Sentiment Analysis (ABSA) aims to determine the sentiment polarity towards an aspect. There always exists severe domain shift between the pretraining and downstream ABSA datasets. We introduce a unified alignment pretraining framework into the vanilla pretrain-finetune pipeline.
  arXiv  Detail & Related papers  (2021-10-26T04:03:45Z)

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.