8/8 ๐ง Can training close this gap?
Fine-tuning helps, but only modestly. Even after training, agents still struggle to update their beliefs from others' behavior and sanctions.
The broader lesson: Learning from one's own mistakes is not enough for normative competence. Building socially intelligent AI may require moving beyond individual trial-and-error toward genuine social learning.
A big thank you to my wonderful collaborators and my advisors @tianminshu@ghadfield for their support throughout this work!
Can agents figure out the social norms of a new communityโwithout being told what they are?
๐ Humans are normatively competent. We watch what others do and what gets corrected, and we adjust our behaviors accordingly.
Introducing NormGym โ a benchmark for evaluating normative competence through social interaction.
๐ค Main takeaway: Agents can sometimes learn from their own norm violations, but struggle to learn from the behavior of others around them. This points to the challenge and the need to build agents that can think and act as members of a community, rather than just pursuing their own goals.
๐ Paper: https://t.co/Mw4kgwMWTf
๐ Project: https://t.co/aNR6cwomWm
7/8 ๐ฅ Humans learn differently.
Human participants reported relying mostly on watching others, or on a combination of observed and firsthand evidence.
That contrast matters: a socially competent agent should be able to learn from what happens to othersโnot need to make every mistake itself.