Henry Xie
Making human-focused AI available on personal devices
Xie is also a three-time semifinalist in the CyberPatriot National Youth Cyber Defense Competition, one of the main youth cybersecurity competitions in the United States, according to background information. Alongside his research, Xie has also worked on technology access and AI ethics in his local community. The son of Fei Xie and Huaiyu Liu, he is president of the computer science club and captain of the varsity swim team at Westview High School. Xie also co-founded Youth for Empathetic AI, a student-led non-profit group that connects high school researchers with academic mentors. The organisation aims to encourage the development of human-focused artificial intelligence and greater involvement by young people in discussions about AI governance.
A 17-year-old student from Portland, Oregon, has developed a two-step training method that helps small artificial intelligence models copy the empathetic communication style of much larger systems. The method achieved higher empathy ratings in at least 90 per cent of comparison tests. Official information published by the Society for Science confirmed that Xie was named a national finalist and received a $25,000 award for his research into ethical artificial intelligence. Evaluation results published by the Society for Science show that when the responses were assessed using LLMs, the small models trained with Xie’s method were rated as more empathetic than basic, untrained SLMs in at least 90 per cent of test situations.
Henry Xie, a senior at Westview High School in Portland, created the computer science project “Distilling Empathy From Large Language Models” for the Regeneron Science Talent Search, the leading science and mathematics competition for high school seniors in the United States. Xie’s method tackles a key problem in current AI technology. Large language models (LLMs) with billions of parameters, such as OpenAI’s ChatGPT and Google’s Gemini, can produce detailed and emotionally supportive responses. However, they need large amounts of computing power, costly server systems and a constant internet connection. Small language models (SLMs), on the other hand, can run directly on devices such as smartphones, laptops and wearable technology. To close this gap, Xie created a two-step training framework that transfers empathetic reasoning from large AI systems to smaller models without increasing their memory or processing requirements. During the first training stage, Xie collected published examples of empathetic conversations produced by LLMs such as ChatGPT and Gemini. He based the data on established psychological ideas about communication and supportive listening. He then used these examples to train small language models and teach them basic patterns for using empathetic language. In the second stage, Xie added specific prompts that encouraged the smaller models to assess their own possible responses. By comparing stronger empathetic answers with weaker or dismissive ones, the models learned to choose responses that better recognised and responded to people’s emotional states.