157 - Diyi Yang: Socially Aware Large Language Models
🎯 Summary
Podcast Episode Summary: 157 - Diyi Yang: Socially Aware Large Language Models
This episode features Professor Diyi Yang discussing her research at the intersection of Computer Science, NLP, and Social Sciences, focusing on developing Socially Aware Language Technologies. The core narrative revolves around moving AI beyond mere technical accuracy to understanding and responding to the complex social signals inherent in human communication.
1. Focus Area
The primary focus is Socially Aware Natural Language Processing (NLP), defined by three dimensions: Social Factors (speaker/receiver relations, context, norms), Social Interactions (governing norms of communication), and Social Implication (the broader societal impact of NLP systems). A major application discussed is using LLMs for Social Skill Training, specifically through the AI Partner AI Mentor (APAM) framework.
2. Key Technical Insights
- APAM Framework for Training: The introduction of the AI Partner (role-playing conversational agent) and AI Mentor (domain-expert coach providing feedback) framework to make social skills training (like conflict resolution) accessible and psychologically safe.
- Addressing Simulation Realism: To overcome LLMs creating caricatures, the approach involves collaborating with domain experts to build typical interaction templates rather than simulating specific individuals, incorporating nuanced social behaviors directly into the simulation design.
- Self-Critique for Contextual Assessment: Techniques like self-critique and self-improvement are used during generation, where the AI Partner assesses its own responses against domain knowledge and appropriateness before outputting them, enhancing realism.
3. Business/Investment Angle
- Transformative Training Market: LLMs offer a scalable, interactive, and personalized solution to training skills that are traditionally time-consuming and expensive (e.g., conflict resolution, counseling).
- Empowering Human Experts: The technology is positioned not as a replacement but as a collaborator (“Helping the Helper”), significantly aiding senior supervisors and instructors by providing personalized, scalable initial feedback to novices.
- Shift from “Book Smart” to “Street Smart”: Demonstrated success in improving interactive performance (conflict resolution) without changing underlying knowledge suggests a high market value for systems that bridge theoretical knowledge with practical application.
4. Notable Companies/People
- Professor Diyi Yang (Stanford): The central expert, leading the Social and Language Technologies Lab, focusing on human-centric NLP.
- Souda Karajah (Stanford): The host and interviewer, highlighting the importance of human elements in AI development.
- Rehearsal System: A specific implementation of the APAM framework used to teach conflict resolution skills, leveraging interest-based negotiation theory.
- CARE System: A system developed to help novice counselors practice empathy and reflection with AI patients, coached by an AI Mentor.
5. Future Implications
The industry is moving toward human-centric AI where social and cultural awareness is integrated into model development and evaluation. Future work aims to incorporate deeper cultural dynamics into the APAM framework. There is also a suggestion to expand training into physical or 3D spaces to incorporate non-verbal cues (posture, eye contact) crucial for complete social skill mastery.
6. Target Audience
AI/ML Researchers focusing on alignment, safety, and human-AI interaction; EdTech Developers looking to integrate advanced personalized training solutions; and Organizational Development/HR Professionals interested in scalable soft-skills training methodologies.
🏢 Companies Mentioned
💬 Key Insights
"the goal of how we build technology is not for the purpose of building technology. The goal is to think about how technology would help us, help more with human touch, help us with developing better and more meaningful interactions with each other."
"I think this is a great direction, and when I think about the skills, I feel like I would encourage people to have a more open mind, understand the mindset about the space. It does not only require social science or social insights, it also requires some kind of computational methods."
"researchers found that large language models tend to produce caricatures when it comes to social simulation, or there may be cultural biases, or they might hallucinate when it comes to specific domains that they don't have a really good knowledge of."
"the second kind of social technical challenge is more about this kind of scenarios that people may practice and then they may develop some kind of reliance there... we worry that one of the issues is this kind of over-reliance."
"We built a system called CARE where novice counselors can actually practice with different types of AI patients. We created those AI patients based on the learning materials that counselors need to learn in their training. And we also built an AI Mentor so that when they practice, they can also get feedback along the way."
"We saw that people's knowledge about conflict resolution actually didn't change at all. However, if you let them do an interactive conflict resolution, people who practiced with our systems actually did much better compared to people who haven't used the system."