Paper · arXiv 2508.07750
Learning to Align, Aligning to Learn: A Unified Approach for Self-Optimized Alignment
Haowen Wang, Yun Yue, Zhiling Ye, Shuowen Zhang, Lei Fan, Jiaxin Liang, Jiadi Jiang, Cheng Wei, Jingyuan Deng, Xudong Han, Ji Li, Chunxiao Guo, Peng Wei, Jian Wang, Jinjie Gu
GRAO, a unified framework combining SFT and RL, enhances language model alignment through multi-sample generation, Group Direct Alignment Loss, and reference-aware parameter updates, demonstrating superior performance across human alignment tasks.
Abstract
Alignment methodologies have emerged as a critical pathway for enhancing language model alignment capabilities. While SFT (supervised fine-tuning) accelerates convergence through direct token-level loss intervention, its efficacy is constrained by offline policy trajectory. In contrast, RL(reinforcement learning) facilitates exploratory policy optimization, but suffers from low sample efficiency and stringent dependency on high-quality base models. To address these dual challenges, we propose GRAO (Group Relative Alignment Optimization), a unified framework that synergizes the respective strengths of SFT and RL through three key innovations: 1) A multi-sample generation strategy enabling comparative quality assessment via reward feedback; 2) A novel Group Direct Alignment Loss formulation leveraging intra-group relative advantage weighting; 3) Reference-aware parameter updates guided by pairwise preference dynamics. Our theoretical analysis establishes GRAO's convergence guarantees and sample efficiency advantages over conventional approaches. Comprehensive evaluations across complex human alignment tasks demonstrate GRAO's superior performance, achieving 57.70\%,17.65\% 7.95\% and 5.18\% relative improvements over SFT, DPO, PPO and GRPO baselines respectively. This work provides both a theoretically grounded alignment framework and empirical evidence for efficient capability evolution in language models.