🖼 🤖Spontaneous Formation of Social Norms in the Group Interaction of Large Language Models

🤖Spontaneous Formation of Social Norms in the Group Interaction of Large Language Models

* A study published in Science Advances shows that large language models (LLMs) can spontaneously form social norms in group interactions, similar to human social behavior.
* The study used Anthropic's Claude model to conduct a naming game experiment. After 24 copies were paired, a unified letter selection tendency gradually formed.
* The experiment still holds when expanded to 200 model copies and 26 letter options, and it was also verified on Meta's Llama model.
* The study found that collective biases develop in the group interaction of models, similar to group behavior in human society, which may bring potential risks.
* When a small number of models preset to always propose new names are introduced into the group, and when they reach a certain number, the existing norms can be overturned and new rules can be established.

(IT Industry Information)

via Teahouse - Telegram Channel