Model Extrapolation Expedites Alignment
This addresses efficiency in aligning LLMs with human preferences, offering a practical solution for researchers and practitioners, though it appears incremental as it builds on existing alignment methods.
The paper tackles the high computational cost of preference alignment training for large language models by proposing ExPO, a method that amplifies parameter changes from partial training to accelerate alignment. Results show ExPO boosts a DPO model trained with only 20% steps to outperform the fully-trained one and improves models on AlpacaEval 2.0 and MT-Bench benchmarks.
Given the high computational cost of preference alignment training of large language models (LLMs), exploring efficient methods to reduce the training overhead remains an important and compelling research problem. Motivated by the observation that alignment training typically involves only small parameter changes without injecting new knowledge into models, we propose a straightforward method called ExPO (model extrapolation) to expedite LLMs' alignment with human preferences. Given a partially-trained model and its initial SFT checkpoint, ExPO improves the implicit optimization objective of alignment training by simply amplifying the parameter change based on a first-order approximation, without any additional training overhead. Through controlled experiments, we demonstrate that ExPO boosts a DPO model trained with only 20% steps to outperform the fully-trained one. Moreover, we show that ExPO notably improves existing open-source LLMs (ranging from 1.8B to 70B parameters) on the leading AlpacaEval 2.0 and MT-Bench benchmarks, which highlights ExPO's broader utility in efficiently enhancing LLM alignment.