Not Everything is All You Need: Toward Low-Redundant Optimization for Large Language Model Alignment
This work addresses alignment inefficiencies in LLMs for AI safety and performance, though it is incremental as it builds on existing parameter-efficient fine-tuning methods.
The paper tackles the problem of large language models overfitting to superficial patterns during alignment with human preferences by proposing ALLO, a method that selectively updates only the most relevant neurons and tokens, resulting in improved convergence and final performance across 10 datasets.
Large language models (LLMs) are still struggling in aligning with human preference in complex tasks and scenarios. They are prone to overfit into the unexpected patterns or superficial styles in the training data. We conduct an empirical study that only selects the top-10\% most updated parameters in LLMs for alignment training, and see improvements in the convergence process and final performance. It indicates the existence of redundant neurons in LLMs for alignment training. To reduce its influence, we propose a low-redundant alignment method named \textbf{ALLO}, focusing on optimizing the most related neurons with the most useful supervised signals. Concretely, we first identify the neurons that are related to the human preference data by a gradient-based strategy, then identify the alignment-related key tokens by reward models for computing loss. Besides, we also decompose the alignment process into the forgetting and learning stages, where we first forget the tokens with unaligned knowledge and then learn aligned knowledge, by updating different ratios of neurons, respectively. Experimental results on 10 datasets have shown the effectiveness of ALLO. Our code and data are available at \url{https://github.com/RUCAIBox/ALLO}.