Tong Zhang

Yong Lin, Hangyu Lin, Wei Xiong et al.

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

Jipeng Zhang, Jianshu Zhang, Yuanzhe Li et al.

Large Language Models (LLMs) demonstrate strong proficiency in generating code for high-resource programming languages (HRPLs) like Python but struggle significantly with low-resource programming languages (LRPLs) such as Racket or D. This performance gap deepens the digital divide, preventing developers using LRPLs from benefiting equally from LLM advancements and reinforcing disparities in innovation within underrepresented programming communities. While generating additional training data for LRPLs is promising, it faces two key challenges: manual annotation is labor-intensive and costly, and LLM-generated LRPL code is often of subpar quality. The underlying cause of this issue is the gap between natural language to programming language gap (NL-PL Gap), which is especially pronounced in LRPLs due to limited aligned data. In this work, we introduce a novel approach called Bridge-Coder, which leverages LLMs' intrinsic capabilities to enhance the performance on LRPLs. Our method consists of two key stages. Bridge Generation, where we create high-quality dataset by utilizing LLMs' general knowledge understanding, proficiency in HRPLs, and in-context learning abilities. Then, we apply the Bridged Alignment, which progressively improves the alignment between NL instructions and LRPLs. Experimental results across multiple LRPLs show that Bridge-Coder significantly enhances model performance, demonstrating the effectiveness and generalization of our approach. Furthermore, we offer a detailed analysis of the key components of our method, providing valuable insights for future work aimed at addressing the challenges associated with LRPLs.