Quan Ze Chen
banner
cqz.bsky.social
Quan Ze Chen
@cqz.bsky.social
Check out our preprint here: arxiv.org/abs/2411.10912
This paper was a collaborative effort from our wonderful team ❤️ @kjfeng.me @chanpark.bsky.social @axz.bsky.social
SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment
When different groups' values differ, one approach to model alignment is to steer models at inference time towards each group's preferences. However, techniques like in-context learning only consider ...
arxiv.org
March 17, 2025 at 5:57 PM
With SPICA, we show the need to not only capture preferences, but also recognize and prioritize norms when it comes to in-context pluralistic alignment.

(8/9)
March 17, 2025 at 5:55 PM
But, more importantly, groups that are often less well represented in alignment datasets see the biggest improvements.

(7/9)
March 17, 2025 at 5:55 PM
Through human evaluations, we find that SPICA-aligned outputs are preferred more on average…

(6/9)
March 17, 2025 at 5:55 PM
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)
March 17, 2025 at 5:54 PM
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)
March 17, 2025 at 5:54 PM
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)
March 17, 2025 at 5:53 PM
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)
March 17, 2025 at 5:53 PM
Reposted by Quan Ze Chen
Next @cqz.bsky.social gave a talk on Wed on targeted interventions to reduce uncertainty in judgments. Paper here: dl.acm.org/doi/10.1145/... He also discussed how it fits into his broader research trajectory and agenda, as he's headed onto the job market this year!
October 19, 2023 at 4:36 AM