contact: [email protected]

Summary

• Generative AI models reproduce human biases in their training data and further amplify them through mechanisms such as mode collapse
• The loss of diversity produces homogenization
• It's hard to capture what diversity is meaningful to different stakeholders.
• This work addresses the homogenization problem by
  • Considering what diversity is meaningful
  • Studying how diversity is mediated by LLMs’ internals (mech interp)
  • Developing ways to promote diversity in LLMs

Theory of change

• Homogenization impoverishes everyone and disproportionately harms the people on the margins by amplifying social biases. Today’s social problems can escalate to catastrophic levels because of GenAI
• Additionally, diversity is at the core of technical problems (like hallucinations and out-of-distribution robustness)
• This work directly addresses the homogenization problem by
  • Studying how diversity is mediated in LLMs’ internals (mech interp)
  • Developing ways to promote diversity in LLMs

The problem

It is hard to adapt and deploy GenAI across diverse settings.

• Generative AI models reproduce human biases