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

arXiv: https://arxiv.org/abs/2504.02193 · HTML: https://mlsanity.com/paper/2504.02193/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 28 of its 37 references that are in the map.
- First posted on arXiv 3 Apr 2025.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)

## Builds on (most cited first, 28 of 28)

- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [TruthfulQA: Measuring How Models Mimic Human Falsehoods](https://mlsanity.com/paper/2109.07958/) (2021)
- [Jailbroken: How Does LLM Safety Training Fail?](https://mlsanity.com/paper/2307.02483/) (2023)
- [Jailbreaking Black Box Large Language Models in Twenty Queries](https://mlsanity.com/paper/2310.08419/) (2023)
- [Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference](https://mlsanity.com/paper/2403.04132/) (2024)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [SimPO: Simple Preference Optimization with a Reference-Free Reward](https://mlsanity.com/paper/2405.14734/) (2024)
- [A General Theoretical Paradigm to Understand Learning from Human Preferences](https://mlsanity.com/paper/2310.12036/) (2023)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [ORPO: Monolithic Preference Optimization without Reference Model](https://mlsanity.com/paper/2403.07691/) (2024)
- [Jailbreaking ChatGPT via Prompt Engineering: An Empirical Study](https://mlsanity.com/paper/2305.13860/) (2023)
- [Safety Alignment Should Be Made More Than Just a Few Tokens Deep](https://mlsanity.com/paper/2406.05946/) (2024)
- [ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection](https://mlsanity.com/paper/2203.09509/) (2022)
- [UltraFeedback: Boosting Language Models with Scaled AI Feedback](https://mlsanity.com/paper/2310.01377/) (2023)
- [Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation](https://mlsanity.com/paper/2401.08417/) (2024)
- [Multi-step Jailbreaking Privacy Attacks on ChatGPT](https://mlsanity.com/paper/2304.05197/) (2023)
- [Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models](https://mlsanity.com/paper/2307.14539/) (2023)
- [Disentangling Length from Quality in Direct Preference Optimization](https://mlsanity.com/paper/2403.19159/) (2024)
- [Direct Nash Optimization: Teaching Language Models to Self-Improve with General Preferences](https://mlsanity.com/paper/2404.03715/) (2024)
- [What is in Your Safe Data? Identifying Benign Data that Breaks Safety](https://mlsanity.com/paper/2404.01099/) (2024)
- [Rule Based Rewards for Language Model Safety](https://mlsanity.com/paper/2411.01111/) (2024)
- [From Lists to Emojis: How Format Bias Affects Model Alignment](https://mlsanity.com/paper/2409.11704/) (2024)

## Built on by (most cited first, 5 of 5)

- [CARE: Multilingual Human Preference Learning for Cultural Awareness](https://mlsanity.com/paper/2504.05154/) (2025)
- [DRIFT: Learning from Abundant User Dissatisfaction in Real-World Preference Learning](https://mlsanity.com/paper/2510.02341/) (2025)
- [Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization](https://mlsanity.com/paper/2509.23371/) (2025)
- [LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X](https://mlsanity.com/paper/2510.13928/) (2025)
- [Smaller Models, Better Rejects: Preference Distillation Scaling](https://mlsanity.com/paper/2609.38987/) (2026)

## Abstract (arXiv)

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.