ARIES: Stimulating Self-Refinement of Large Language Models by Iterative Preference Optimization

A truly intelligent Large Language Model (LLM) should be capable of correcting errors in its responses through external interactions. However, even the most advanced models often face challenges in improving their outputs. In this paper, we explore how to cultivate LLMs with the self-refinement capability through iterative preference training, and how this ability can be leveraged to improve model performance during inference. To this end, we introduce a novel post-training and inference framework, called ARIES: Adaptive Refinement and Iterative Enhancement Structure. This method iteratively performs preference training and self-refinement-based data collection. During training, ARIES strengthen the model's direct question-answering capability while simultaneously unlocking its self-refinement potential. During inference, ARIES harnesses this self-refinement capability to generate a series of progressively refined responses, which are then filtered using either the Reward Model Scoring or a simple yet effective Rule-Based Selection mechanism, specifically tailored to our approach, to construct a dataset for the next round of preference training. Experimental results demonstrate the remarkable performance of ARIES. When applied to the Llama-3.1-8B model and under the self-refinement setting, ARIES surpasses powerful models such as GPT-4o, achieving 62.3% length-controlled (LC) and a 63.3% raw win rates on AlpacaEval 2, outperforming Iterative DPO by 27.8% and 35.5% respectively, as well as a 50.3% win rate on Arena-Hard, surpassing Iterative DPO by 26.6%. Furthermore, ARIES consistently enhances performance on mathematical reasoning tasks like GSM8K and MATH.

Token-level Direct Preference Optimization

Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous methods, which face challenges in divergence efficiency, TDPO incorporates forward KL divergence constraints for each token, improving alignment and diversity. Utilizing the Bradley-Terry model for a token-based reward system, TDPO enhances the regulation of KL divergence, while preserving simplicity without the need for explicit reward modeling. Experimental results across various text tasks demonstrate TDPO's superior performance in balancing alignment with generation diversity. Notably, fine-tuning with TDPO strikes a better balance than DPO in the controlled sentiment generation and single-turn dialogue datasets, and significantly improves the quality of generated responses compared to both DPO and PPO-based RLHF methods. Our code is open-sourced at https://github.com/Vance0124/Token-level-Direct-Preference-Optimization.