But, more importantly, groups that are often less well represented in alignment datasets see the biggest improvements.

(7/9)

Through human evaluations, we find that SPICA-aligned outputs are preferred more on average…

(6/9)

We then make use of these metrics during the retrieval process, producing pluralistically aligned examples that both reflect group preferences, and also their norms.

(5/9)

In SPICA, we sample **individual preferences** of members in a group to create metrics inspired by social norm theory that inform us of how each group prioritizes which examples they care more about (best illustrates group norms)

(4/9)

We argue that group level differences extend beyond their preferences for how to answer, and that different groups can also have preferences around which queries are better examples of how they prioritize their values.

(3/9)

Traditional in-context alignment (ICA) retrieves demonstration examples (query & answer) by finding those most similar to a new query. However, when there is a plurality of groups to align to, the same queries get picked regardless of group.

(2/9)

In-context learning can be an effective way to conduct value alignment of LLMs through examples, but when there are multiple pluralistic groups, are the best examples for one group also the ones for another?

We explore this in our paper 🌟SPICA🌟
(🧵1/9)