Direct Preference Optimization (DPO) has emerged as a simple and effective approach for aligning large language models (LLMs) with human preferences, bypassing the need for a learned reward model. Despite its growing adoption, a fundamental question remains open: what characteristics of preference data are most critical for DPO performance? In this work, we provide a systematic study of how preference data distribution influences DPO, from both theoretical and empirical perspectives. We show that the quality of chosen responses plays a dominant role in optimizing the DPO objective, while the quality of rejected responses may have relatively limited impact. Our theoretical analysis characterizes the optimal response distribution under DPO and reveals how contrastiveness between responses helps primarily by improving the chosen samples. We further study an online DPO setting and show it effectively reduces to supervised fine-tuning on the chosen responses. Extensive experiments across diverse tasks confirm our findings: improving the quality of chosen responses consistently boosts performance regardless of the quality of the rejected responses. We also investigate the benefit of mixing the on-policy data. Our results interpret the mechanism behind some widely adopted strategies and offer practical insights for constructing highimpact preference datasets for LLM alignment.

Direct Preference Optimization (DPO) has emerged as a simple and effective approach for aligning large language models (LLMs) with human preferences, bypassing the need for a learned reward model. Despite its growing adoption, a fundamental question remains open: what characteristics of preference data are most critical for DPO performance? In this work, we provide a systematic study of how preference data distribution influences DPO, from both theoretical and empirical perspectives. We show that the quality of chosen responses plays a dominant role in optimizing the DPO objective, while the quality of rejected responses may have relatively limited impact. Our theoretical analysis characterizes the optimal response distribution under DPO and reveals how contrastiveness between responses helps primarily by improving the chosen samples. We further study an online DPO setting and show it effectively reduces to supervised fine-tuning on the chosen responses. Extensive experiments across diverse tasks confirm our findings: improving the quality of chosen responses consistently boosts performance regardless of the quality of the rejected responses. We also investigate the benefit of mixing the on-policy data. Our results interpret the mechanism behind some widely adopted strategies and offer practical insights for constructing high-impact preference datasets for LLM alignment.

Large Language Models (LLMs) as autonomous agents are increasingly tasked with solving complex, long-horizon problems. Aligning these agents via preference-based offline methods like Direct Preference Optimization (DPO) is a promising direction, yet it faces a critical granularity mismatch. Trajectory-level DPO provides a signal that is too coarse for precise credit assignment, while step-level DPO is often too myopic to capture the value of multi-step behaviors. To resolve this challenge, we introduce Hierarchical Preference Learning (HPL), a hierarchical framework that optimizes LLM agents by leveraging preference signals at multiple, synergistic granularities. While HPL incorporates trajectory- and step-level DPO for global and local policy stability, its core innovation lies in group-level preference optimization guided by a dual-layer curriculum. Our approach first decomposes expert trajectories into semantically coherent action groups and then generates contrasting suboptimal groups to enable preference learning at a fine-grained, sub-task level. Then, instead of treating all preference pairs equally, HPL introduces a curriculum scheduler that organizes the learning process from simple to complex. This curriculum is structured along two axes: the group length, representing sub-task complexity, and the sample difficulty, defined by the reward gap between preferred and dispreferred action groups. Experiments on three challenging agent benchmarks show that HPL outperforms existing state-of-the-art methods. Our analyses demonstrate that the hierarchical DPO loss effectively integrates preference signals across multiple granularities, while the dual-layer curriculum is crucial for enabling the agent to solve a wide range of tasks, from simple behaviors to complex multi-step sequences.

The code generation capabilities of Large Language Models (LLMs) have advanced applications like tool invocation and problem-solving. However, improving performance in code-related tasks remains challenging due to limited training data that is verifiable with accurate test cases. While Direct Preference Optimization (DPO) has shown promise, existing methods for generating test cases still face limitations. In this paper, we propose a novel approach that splits code snippets into smaller, granular blocks, creating more diverse DPO pairs from the same test cases. Additionally, we introduce the Abstract Syntax Tree (AST) splitting and curriculum training method to enhance the DPO training. Our approach demonstrates significant improvements in code generation tasks, as validated by experiments on benchmark datasets such as HumanEval (+), MBPP (+), APPS, LiveCodeBench, and BigCodeBench. Code and data are available at https://github.com/SenseLLM/StructureCoder.

Direct Preference Optimization (DPO) is a simple and efficient framework that has attracted substantial attention. However, it often struggles to meet its primary objectives -- increasing the generation probability of chosen responses while reducing that of rejected responses -- due to the dominant influence of rejected responses on the loss function. This imbalance leads to suboptimal performance in promoting preferred responses. In this work, we systematically analyze the limitations of DPO and existing algorithms designed to achieve the objectives stated above. To address these limitations, we propose Bounded-DPO (BDPO), a novel method that bounds the influence of rejected responses while maintaining the original optimization structure of DPO. Through theoretical analysis and empirical evaluations, we demonstrate that BDPO achieves a balanced optimization of the chosen and rejected responses, outperforming existing algorithms.

Direct Preference Optimization (DPO) often struggles with long-chain mathematical reasoning. Existing approaches, such as Step-DPO, typically improve this by focusing on the first erroneous step in the reasoning chain. However, they overlook all other steps and rely heavily on humans or GPT-4 to identify erroneous steps. To address these issues, we propose Full-Step-DPO, a novel DPO framework tailored for mathematical reasoning. Instead of optimizing only the first erroneous step, it leverages step-wise rewards from the entire reasoning chain. This is achieved by training a self-supervised process reward model, which automatically scores each step, providing rewards while avoiding reliance on external signals. Furthermore, we introduce a novel step-wise DPO loss, which dynamically updates gradients based on these step-wise rewards. This endows stronger reasoning capabilities to language models. Extensive evaluations on both in-domain and out-of-domain mathematical reasoning benchmarks across various base language models, demonstrate that Full-Step-DPO achieves superior performance compared to state-of-the-art baselines.

Preference learning, or the task of aligning generative models to preference comparison data, has yet to reach the conceptual maturity of classification, density estimation, etc. To close this gap, this work presents a framework to understand preference learning starting from the sampling distribution of pairwise preference data. First, we prove that the only evaluation of a generative model that respects both preferences and prevalences in the data distribution is a form of win rate, justifying win rate as the focal point to understand preference learning. We then analyze preference learning methods as win rate optimization (WRO) or non-WRO. We present novel instances of WRO beyond existing examples (RLHF, NLHF) and identify two key theoretical benefits of all such methods. We prove that common non-WRO methods like DPO and SFT on preferred samples lack these properties and suggest ways to mitigate such theoretical limitations. We also show that WRO underperforms in practice due optimization difficulties and that optimization success predicts performance better than choices which affect the objective's solution. Our analysis highlights best practices for existing methods and provides recommendations for future research, guided by the principle that one should either align non-WRO methods more closely with WRO or improve the optimization of WRO objectives.

Aligning Large Language Models (LLMs) to human preferences in content, style, and presentation is challenging, in part because preferences are varied, contextdependent, and sometimes inherently ambiguous. While successful, Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) are prone to the issue of proxy reward overoptimization. Analysis of the DPO loss reveals a critical need for regularization for mislabeled or ambiguous preference pairs to avoid reward hacking. In this work, we develop a pessimistic framework for DPO by introducing preference uncertainty penalization schemes, inspired by offline reinforcement learning. The penalization serves as a correction to the loss which attenuates the loss gradient for uncertain samples. Evaluation of the methods is performed with GPT2 Medium on the Anthropic-HH dataset using a model ensemble to obtain uncertainty estimates, and shows improved overall performance compared to vanilla DPO, as well as better completions on prompts from high-uncertainty chosen/rejected responses. Aligning LLMs to human preferences in content, style, and presentation has become a central challenge in improving and deploying LLMs, leading to the advent of Reinforcement Learning with Human Feedback (RLHF), now a prominent technique to fine-tune state-of-the-art LLMs (Casper et al., 2023). The standard RLHF pipeline involves human feedback collection, reward model training, and LLM policy optimization via reinforcement learning (RL). Despite its success, each stage presents challenges, from feedback interpretation and policy generalization to challenging RL implementation (Casper et al., 2023). Direct Preference Optimisation (DPO) (Rafailov et al., 2023) effectively bypasses the reward model by fine-tuning the policy to maximize the likelihood of the preference data under the Bradley-Terry model (A. DPO is easier to implement than RL algorithms, and benefits from computational efficiency and stability by avoiding potential inaccuracies and biases of a reward model (Xu et al., 2024; Casper et al., 2023).

We propose the intermediate direct preference optimization (DPO) method to calculate the DPO loss at selected intermediate layers as an auxiliary loss for finetuning large language models (LLMs). The conventional DPO method fine-tunes a supervised fine-tuning (SFT) model by calculating the DPO loss using logits from the final layer. In our intermediate DPO approach, DPO losses are calculated using the logits from K-selected intermediate layers and averaged to obtain the intermediate DPO loss. For training the intermediate DPO model, the final loss is obtained by calculating the weighted sum of the DPO and intermediate DPO losses. During inference, the intermediate DPO model decodes using the final layer logits similarly to the conventional DPO model. In experiments using the ultrafeedback dataset, the performance of the intermediate DPO model was evaluated using GPT-4. As a result, the intermediate DPO model trained using the intermediate DPO loss calculated at the 22nd layer of a 32-layer SFT model achieved win rates of 52.5% and 67.5% against the conventional DPO and SFT models, respectively, demonstrating the effectiveness of the proposed method. Furthermore, we report the relationships among the position of the selected intermediate layers, the number of layers, and performance.

Large language models in the past have typically relied on some form of reinforcement learning with human feedback (RLHF) to better align model responses with human preferences. However, because of oft-observed instabilities when implementing these RLHF pipelines, various reparameterization techniques have recently been introduced to sidestep the need for separately learning an RL reward model. Instead, directly fine-tuning for human preferences is achieved via the minimization of a single closed-form training objective, a process originally referred to as direct preference optimization (DPO) and followed by several notable descendants. Although effective in certain real-world settings, we introduce new evaluation criteria that serve to highlight unresolved shortcomings in the ability of existing DPO methods to interpolate between a pre-trained reference model and empirical measures of human preferences, as well as unavoidable trade-offs in how low- and high-quality responses are regularized and constraints are handled. Our insights then motivate an alternative DPO-like loss that provably mitigates these limitations. Empirical results serve to corroborate notable aspects of our analyses.