More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment

• URL: http://arxiv.org/abs/2504.02193v1
• Date: Thu, 03 Apr 2025 00:36:40 GMT
• Title: More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment
• Authors: Yifan Wang, Runjin Chen, Bolian Li, David Cho, Yihe Deng, Ruqi Zhang, Tianlong Chen, Zhangyang Wang, Ananth Grama, Junyuan Hong,
• Abstract summary: Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback.<n>Our study reveals a striking, safety-specific phenomenon associated with DPO alignment.<n>Using solely self-generated responses for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models.
• Score: 80.04449725137177
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogrande) by providing diverse responses, it also tends to facilitate reward hacking during training. This can lead to a high attack success rate (ASR) when models encounter jailbreaking prompts. The issue is particularly pronounced when employing stronger models like GPT-4o or larger models in the same family to generate chosen responses paired with target model self-generated rejected responses, resulting in dramatically poorer safety outcomes. Furthermore, with respect to safety, using solely self-generated responses (single-model generation) for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models, whether used directly as chosen data or as part of a multi-model response pool. We demonstrate that multi-model preference data exhibits high linear separability between chosen and rejected responses, which allows models to exploit superficial cues rather than internalizing robust safety constraints. Our experiments, conducted on models from the Llama, Mistral, and Qwen families, consistently validate these findings.

Related papers

• Few-shot Steerable Alignment: Adapting Rewards and LLM Policies with Neural Processes [50.544186914115045]
  Large language models (LLMs) are increasingly embedded in everyday applications. Ensuring their alignment with the diverse preferences of individual users has become a critical challenge. We present a novel framework for few-shot steerable alignment.
  arXiv  Detail & Related papers  (2024-12-18T16:14:59Z)
• Reward-Augmented Data Enhances Direct Preference Alignment of LLMs [63.32585910975191]
  We introduce reward-conditioned Large Language Models (LLMs) that learn from the entire spectrum of response quality within the dataset. We show that our approach consistently boosts DPO by a considerable margin. Our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere data expansion.
  arXiv  Detail & Related papers  (2024-10-10T16:01:51Z)
• Progressively Label Enhancement for Large Language Model Alignment [42.01694160556464]
  Large Language Models (LLM) alignment aims to prevent models from producing content that misaligns with human expectations. We propose PLE, a framework that dynamically adjusts the model's training process based on the evolving quality of the generated data.
  arXiv  Detail & Related papers  (2024-08-05T16:21:17Z)
• Online Self-Preferring Language Models [34.22412851864247]
  Online Self-Preferring (OSP) language models learn from self-generated response pairs and self-judged preference strengths. OSP achieves state-of-the-art alignment performance across various metrics in two widely used human preference datasets.
  arXiv  Detail & Related papers  (2024-05-23T02:13:34Z)
• West-of-N: Synthetic Preferences for Self-Improving Reward Models [20.643537269666137]
  We present a novel approach to improve reward model quality by generating synthetic preference data. We find that this approach improves the performance of any reward model, with an effect comparable to the addition of a similar quantity of human preference data.
  arXiv  Detail & Related papers  (2024-01-22T16:24:43Z)
• Secrets of RLHF in Large Language Models Part II: Reward Modeling [134.97964938009588]
  We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset. We also introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses.
  arXiv  Detail & Related papers  (2024-01-11T17:56:59Z)
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model [119.65409513119963]
  We introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods.
  arXiv  Detail & Related papers  (2023-05-29T17:57:46Z)
• On the Efficacy of Adversarial Data Collection for Question Answering: Results from a Large-Scale Randomized Study [65.17429512679695]
  In adversarial data collection (ADC), a human workforce interacts with a model in real time, attempting to produce examples that elicit incorrect predictions. Despite ADC's intuitive appeal, it remains unclear when training on adversarial datasets produces more robust models.
  arXiv  Detail & Related papers  (2021-06-02T00:48:33Z)

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.