Token-Importance Guided Direct Preference Optimization

Aligning Large Language Models (LLMs) with human preferences is crucial for safe and effective AI interactions. While popular methods like Direct Preference Optimization (DPO) have simplified alignment, they remain sensitive to data noise and overlook the differential importance of individual tokens. Existing token-level approaches often rely on probability prediction or simplistic weighting schemes to obtain token importance, which still cannot fully address these issues. To solve this problem, we propose the Token-Importance Guided Direct Preference Optimization (TI-DPO), a framework that achieves fine-grained semantic control through two synergistic innovations. First, we propose a novel hybrid weighting mechanism that combines gradient attribution with a Gaussian prior, ensuring both the accuracy and robustness of token importance scores. Second, we employ a triplet loss to provide structured guidance for the optimization, explicitly guiding model outputs to approach preferred responses and diverge from non-preferred ones. Experimental results show that TI-DPO achieves higher accuracy and stronger generative diversity, providing more stable and computationally efficient solutions compared with DPO and other RLHF methods. Large Language Models (LLMs) have shown proficiency in Natural Language Processing (NLP) (Gao et al., 2025), logical reasoning (Xie et al., 2025a), and code generation (Xu et al., 2025), emerging as a focal point of recent research. However, as models may generate outputs inconsistent with intended purposes or ethical standards, human preference alignment aims to ensure that LLMs adhere to human values (Liu et al., 2023), producing beneficial and harmless content. Against this backdrop, Reinforcement Learning from Human Feedback (RLHF) has become a prevailing approach for achieving alignment (Hong et al., 2024; Hu et al., 2025).