AlphaPO: Reward Shape Matters for LLM Alignment
This work addresses a specific bottleneck in aligning large language models with human preferences, offering incremental improvements over existing methods.
The paper tackles the problem of likelihood displacement in Direct Alignment Algorithms (DAAs) for LLM alignment by introducing AlphaPO, a method that adjusts the reward function shape using an α-parameter, resulting in 7-10% relative improvement over SimPO and 15-50% over DPO on models like Mistral-7B and Llama3-8B.
Reinforcement Learning with Human Feedback (RLHF) and its variants have made huge strides toward the effective alignment of large language models (LLMs) to follow instructions and reflect human values. More recently, Direct Alignment Algorithms (DAAs) have emerged in which the reward modeling stage of RLHF is skipped by characterizing the reward directly as a function of the policy being learned. Some popular examples of DAAs include Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO). These methods often suffer from likelihood displacement, a phenomenon by which the probabilities of preferred responses are often reduced undesirably. In this paper, we argue that, for DAAs the reward (function) shape matters. We introduce \textbf{AlphaPO}, a new DAA method that leverages an $α$-parameter to help change the shape of the reward function beyond the standard log reward. AlphaPO helps maintain fine-grained control over likelihood displacement and over-optimization. Compared to SimPO, one of the best performing DAAs, AlphaPO leads to about 7\% to 10\% relative improvement in alignment performance for the instruct versions of Mistral-7B and Llama3-8B while achieving 15\% to 50\% relative improvement over DPO on the same models. The analysis and results presented highlight the importance of the reward shape and how one can systematically change it to affect training dynamics, as well as improve alignment performance.