Diversity First, Quality Later: A Two-Stage Assumption for Language Model Alignment

• URL: http://arxiv.org/abs/2508.10530v1
• Date: Thu, 14 Aug 2025 11:05:18 GMT
• Title: Diversity First, Quality Later: A Two-Stage Assumption for Language Model Alignment
• Authors: Zetian Sun, Dongfang Li, Baotian Hu,
• Abstract summary: The alignment of language models (LMs) with human preferences is critical for building reliable AI systems.<n>Recently, Direct Preference Optimization (DPO) was proposed as a LM alignment method that directly optimize the policy from static preference data.<n>We show on-policy data is not always optimal, with systematic effectiveness difference emerging between static and on-policy preference candidates.
• Score: 16.059172179404467
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: The alignment of language models (LMs) with human preferences is critical for building reliable AI systems. The problem is typically framed as optimizing an LM policy to maximize the expected reward that reflects human preferences. Recently, Direct Preference Optimization (DPO) was proposed as a LM alignment method that directly optimize the policy from static preference data, and further improved by incorporating on-policy sampling (i.e., preference candidates generated during the training loop) for better LM alignment. However, we show on-policy data is not always optimal, with systematic effectiveness difference emerging between static and on-policy preference candidates. For example, on-policy data can result in a 3$\times$ effectiveness compared with static data for Llama-3, and a 0.4$\times$ effectiveness for Zephyr. To explain the phenomenon, we propose the alignment stage assumption, which divides the alignment process into two distinct stages: the preference injection stage, which benefits from diverse data, and the preference fine-tuning stage, which favors high-quality data. Through theoretical and empirical analysis, we characterize these stages and propose an effective algorithm to identify the boundaries between them. We perform experiments on 5 models (Llama, Zephyr, Phi-2, Qwen, Pythia) and 2 alignment methods (DPO, SLiC-HF) to show the generalizability of alignment stage assumption and boundary measurement.

Related papers

• How Sampling Shapes LLM Alignment: From One-Shot Optima to Iterative Dynamics [65.67654005892469]
  We show that proper instance-dependent sampling can yield stronger ranking guarantees, while skewed on-policy sampling can induce excessive concentration under structured preferences.<n>We then analyze iterative alignment dynamics in which the learned policy feeds back into future sampling and reference policies.<n>Our theoretical insights extend to Direct Preference Optimization, indicating the phenomena we captured are common to a broader class of preference-alignment methods.
  arXiv  Detail & Related papers  (2026-02-12T17:11:08Z)
• Latent Collective Preference Optimization: A General Framework for Robust LLM Alignment [7.1259212876994695]
  We introduce Latent Collective Preference Optimization (LCPO) to learn the latent collective consensus from noisy data.<n>Our experiments demonstrate LCPO's effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms.<n>When applied to Mistral and Llama 3 models, LCPO-enhanced methods achieve substantial win rate gains on AlpacaEval 2 and Arena-Hard, with improvements of up to 7.0% on both benchmarks.
  arXiv  Detail & Related papers  (2025-09-29T01:17:49Z)
• Beyond Single: A Data Selection Principle for LLM Alignment via Fine-Grained Preference Signals [46.58760908162995]
  We propose a novel, theoretically-grounded data selection principle for large language models.<n>We prove the optimality of this strategy by analyzing the loss bounds of the Direct Preference Optimization objective.<n>Our strategy achieves over 10% relative improvement against both the standard holistic preference and a stronger oracle.
  arXiv  Detail & Related papers  (2025-08-11T05:43:02Z)
• PIPA: Preference Alignment as Prior-Informed Statistical Estimation [57.24096291517857]
  We introduce Pior-Informed Preference Alignment (PIPA), a unified, RL-free probabilistic framework.<n> PIPA accommodates both paired and unpaired data, as well as answer and step-level annotations.<n>By integrating different types of prior information, we developed two variations of PIPA: PIPA-M and PIPA-N.
  arXiv  Detail & Related papers  (2025-02-09T04:31:30Z)
• Self-Steering Optimization: Autonomous Preference Optimization for Large Language Models [79.84205827056907]
  We present Self-Steering Optimization ($SSO$), an algorithm that autonomously generates high-quality preference data.<n>$SSO$ employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data.<n>Our evaluation shows that $SSO$ consistently outperforms baselines in human preference alignment and reward optimization.
  arXiv  Detail & Related papers  (2024-10-22T16:04:03Z)
• 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.<n>We show that our approach consistently boosts DPO by a considerable margin.<n>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)
• A Deep Dive into the Trade-Offs of Parameter-Efficient Preference Alignment Techniques [63.10251271444959]
  Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. We conduct an in-depth investigation of the impact of popular choices for three crucial axes. Our setup spanning over 300 experiments reveals consistent trends and unexpected findings.
  arXiv  Detail & Related papers  (2024-06-07T12:25:51Z)
• 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)
• Statistical Rejection Sampling Improves Preference Optimization [42.57245965632205]
  We introduce a novel approach to source preference data from the target optimal policy using rejection sampling. We also propose a unified framework that enhances the loss functions used in both Sequence Likelihood (SLiC) and Direct Preference Optimization (DPO) from a preference modeling standpoint.
  arXiv  Detail & Related papers  (2023-09-13T01:07:25Z)

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.