To help users better understand the value alignment performance of different large language models, Microsoft Research Asia has launched the Value Compass Benchmarks website, based on its latest value evaluation framework. This framework is designed according to social science theories and can dynamically, scientifically, systematically, and comprehensively assess the values of large models in step with their rapid development. With the Value Compass Benchmarks website, users can conduct in-depth value analysis of mainstream models on the market, enabling them to select the most suitable model for their specific scenarios and address various challenges more efficiently.

At a recent TEDxBeijing Live event, Dr. Xiao-Yuan Yi, a Senior Researcher at Microsoft Research Asia, elucidated the relationship between the values of artificial intelligence (AI) and its behavior. He further delved into methods for effectively and stably aligning AI with human values. Additionally, he presented the latest research progress of the “Value Compass” project at Microsoft Research Asia, which aims to address the challenges of defining, evaluating, and aligning AI’s values. The project is dedicated to the social responsibility of guiding technological innovation with human values, thus fostering a future society where humans and machines coexist harmoniously.

With rapid advancements of artificial intelligence (AI) technology and its capabilities, large models have gradually been integrated into daily life. However, this progress also brings new potential risks, further highlighting the urgency of aligning large models with human values. But which values should AI align with, and how should this alignment be achieved? These questions remain unanswered.

To address these challenges, Microsoft Research Asia has launched the “Value Compass” project, taking an interdisciplinary approach and drawing extensively on theories from ethics and sociology to tackle the issues of defining, evaluating, and aligning values. This article will provide an in-depth analysis of the current state of value alignment in large models and present the latest research achievements from Microsoft Research Asia in this field – the BaseAlign alignment algorithm, based on the Schwartz theory of basic human values.

Recently, the Association for Computing Machinery (ACM), the world’s most influential professional academic organization in the field of computing, announced the 2023 list of ACM Fellows. Dr. Xing Xie, a Senior Principal Researcher at Microsoft Research Asia, was named a Fellow for his outstanding contributions to spatial data mining and recommender systems.

The rapid development of artificial intelligence (AI) is having an ever-increasing impact on human society. To ensure AI becomes a technology that is socially responsible, we have chosen “Societal AI” as our research direction, collaborating across disciplines with social sciences such as psychology, sociology, and law. Our goal is to explore how AI can understand and adhere to the mainstream values of human society, make decisions that align with human expectations, and enable people to accurately assess the true value orientations and intelligence levels of AI through more robust evaluation models.

——Xing Xie
Senior Principal Researcher, Microsoft Research Asia

Xing Xie is a Senior Principal Researcher at Microsoft Research Asia and a pioneer in the field of spatial data mining. Dedicated to ensuring the responsible and value-aligned use of artificial intelligence, his team focuses on the research direction of “Societal AI”. This approach aims to guarantee the fairness, reliability and safety, privacy protection, inclusivity, transparency, and accountability of technology and large model applications.

Their efforts include designing effective privacy protection methods to safeguard data used for training or fine-tuning large models and developing easy-to-understand explanation methods to help users understand how large models operate. They also work on mitigating or eliminating inherent social biases and harmful content in training data, making large models safer and more robust in dynamic environments.