Alexandria: A Library of Pluralistic Values for Realtime Re-Ranking of Social Media Feeds

TL;DR

Engagement-focused ranking often fails to capture broader societal values. Alexandria introduces a library of 78 post-level values derived from six value systems and operationalized with an LLM labeling pipeline to enable real-time re-ranking on X/Twitter via a Chrome extension, with ranking scores computed as $s_i = \\mathbf{r}_i \\cdot \\mathbf{w}$. Through qualitative (N=12) and quantitative (N=257) studies, the work demonstrates improved value alignment and user control when using the full library versus single-value baselines. These findings support scalable, user-centric design of pluralistic feeds and offer governance considerations for a marketplace of values within platform design.

Abstract

Social media feed ranking algorithms fail when they too narrowly focus on engagement as their objective. The literature has asserted a wide variety of values that these algorithms should account for as well -- ranging from well-being to productive discourse -- far more than can be encapsulated by a single topic or theory. In response, we present a $\textit{library of values}$ for social media algorithms: a pluralistic set of 78 values as articulated across the literature, implemented into LLM-powered content classifiers that can be installed individually or in combination for real-time re-ranking of social media feeds. We investigate this approach by developing a browser extension, $\textit{Alexandria}$, that re-ranks the X/Twitter feed in real time based on the user's desired values. Through two user studies, both qualitative (N=12) and quantitative (N=257), we found that diverse user needs require a large library of values, enabling more nuanced preferences and greater user control. With this work, we argue that the values criticized as missing from social media ranking algorithms can be operationalized and deployed today through end-user tools.

Figures (4)

• Figure 1: After selecting the values they wish to see from an initial set of five, users can then iteratively refine their feed by upranking or downranking additional values. The Onboarding Page [fill color=black,inner color=white,]A [fill color=black,inner color=white,]A is shown to users when they first install the extension. At any time, users can refine the values they selected using the Value Controls Page [fill color=black,inner color=white,]B [fill color=black,inner color=white,]B .
• Figure 2: Twenty-three of the 78 values in the full library are significantly different ($p < 0.05$) from zero. We visualize the mean weights, which can range from -1 to 1, applied to each value by participants who had access to the full library of values. For each value, we perform a one-sample t-test with a hypothetical mean of 0. We then apply the Benjamini-Hochberg procedure on the $p$-values. Error bars represent 95% confidence intervals. $^* p<0.05$. $^{**} p<0.01$. $^{***} p<0.001$.
• Figure A.1: Our value library begins with a set of values for social media algorithms drawn from existing literature and frameworks. We adapt these frameworks' terms and definitions, then use a large language model (LLM) pipeline to annotate expressions of those values in social media content, creating algorithmic operationalizations of each value.
• Figure A.2: Distribution of number of values from our library found on Twitter posts. Less than 3% of posts have no values present.