# From Instructions to Intrinsic Human Values -- A Survey of Alignment Goals for Big Models

**Authors:** Jing Yao, Xiaoyuan Yi, Xiting Wang, Jindong Wang, Xing Xie  
**arXiv:** [2308.12014](https://arxiv.org/abs/2308.12014)  
**Date:** 2023-08-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.12014v1/

## Key points

- Big models, specifically Large Language Models (LLMs), pose challenges and potential risks
- Integration of LLMs into human lives raises concerns about social harm
- Efforts have been made to align LLMs with humans to follow instructions and satisfy preferences
- The question of "what to align with" has not been fully explored
- Inappropriate alignment goals may have negative consequences
- This paper provides a comprehensive survey of alignment goals proposed in existing work
- Alignment goals are categorized into fundamental abilities, task-specific objectives, and value orientation
- Alignment goals have evolved from focusing on fundamental abilities to prioritizing intrinsic human values
- Challenges in achieving intrinsic value alignment between big models and humans are discussed
- Available resources for future research on aligning big models with human values are provided

## Summary

In the paper titled "From Instructions to Intrinsic Human Values - A Survey of Alignment Goals for Big Models," authors Jing Yao, Xiaoyuan Yi, Xiting Wang, Jindong Wang, and Xing Xie discuss the challenges and potential risks associated with big models, specifically Large Language Models (LLMs). These models are pre-trained on massive amounts of data and consist of a large number of parameters. While they offer improved performance across various tasks and possess emergent capabilities not found in smaller models, their integration into everyday human lives raises concerns about potential social harm. To address these concerns, efforts have been made to align LLMs with humans in order to better follow user instructions and satisfy human preferences. However, the question of "what to align with" has not been fully explored, and inappropriate alignment goals may even have negative consequences. Therefore, this paper aims to provide a comprehensive survey of different alignment goals proposed in existing work and trace their evolution paths to identify the most essential goal. The authors approach this investigation from two perspectives: the definition of alignment goals and alignment evaluation. They analyze existing works and categorize alignment goals into three distinct levels: fundamental abilities, task-specific objectives, and value orientation. Through this analysis, they observe a transformation in alignment goals from focusing on fundamental abilities to prioritizing intrinsic human values as the ultimate goal for enhanced LLMs. Based on their findings, the authors further discuss the challenges involved in achieving intrinsic value alignment between big models and humans. They also provide a collection of available resources for future research on aligning big models with human values. Overall, this paper sheds light on the importance of aligning big models with humans while highlighting the need for careful consideration when determining alignment goals. By emphasizing intrinsic human values as a crucial aspect of alignment it offers valuable insights for researchers working towards enhancing LLMs while minimizing potential social harms.