Paper Skygest: Personalized Academic Recommendations on Bluesky

TL;DR

Paper Skygest demonstrates that a personalized academic feed deployed by researchers on Bluesky can achieve sustained organic usage and shift engagement toward scientific content. The authors present a three-component architecture (feed generator endpoint, firehose, and generation module) and provide open-source code to help others reproduce. They show that ranking effects exist in the feed and that adoption increases paper-post interactions, while also enabling large-scale randomized experimental studies without platform partnerships. Finally, they argue that custom feeds enable studying algorithm design, user agency, and new research opportunities on decentralized social platforms.

Abstract

We build, deploy, and evaluate Paper Skygest, a custom personalized social feed for scientific content posted by a user's network on Bluesky and the AT Protocol. We leverage a new capability on emerging decentralized social media platforms: the ability for anyone to build and deploy feeds for other users, to use just as they would a native platform-built feed. To our knowledge, Paper Skygest is the first and largest such continuously deployed personalized social media feed by academics, with over 50,000 weekly uses by over 1,000 daily active users, all organically acquired. First, we quantitatively and qualitatively evaluate Paper Skygest usage, showing that it has sustained usage and satisfies users; we further show adoption of Paper Skygest increases a user's interactions with posts about research, and how interaction rates change as a function of post order. Second, we share our full code and describe our system architecture, to support other academics in building and deploying such feeds sustainably. Third, we overview the potential of custom feeds such as Paper Skygest for studying algorithm designs, building for user agency, and running recommender system experiments with organic users without partnering with a centralized platform.

Figures (5)

• Figure 1: (a) A user's perspective of using Paper Skygest: the user switches to the feed and sees a list of posts about research papers (including the broader thread) by people in their following network. (b) A diagram of the Paper Skygest backend, which is comprised of three main components: (1) the feed generator endpoint, which serves cached recommendations to users, (2) the firehose, which listens for and stores paper posts, and (3) the recommendation generation module, which precomputes the paper posts to show to each Paper Skygest user at regular intervals and caches them.
• Figure 2: Paper Skygest usage data over time. (a) Figure \ref{['fig:usages:users']} shows, for each day since launching, the number of unique accounts that used Paper Skygest and the number of sessions on Paper Skygest (where a session is defined as a request from Bluesky for the first page of posts). The number of unique users has increased over time, while the daily number of sessions has remained stable. Both user and session counts vary throughout the week, with high counts on weekdays and low counts on weekends. (b) Figure \ref{['fig:stickiness']} shows average usage trajectories for users over time, where users are grouped by their percentile of total Paper Skygest accesses. We consider usage data between May 21, 2025 and September 16, 2025, for users who accessed Paper Skygest for the first time prior to May 21, 2025; we average across users and report $95\%$ confidence intervals.
• Figure 3: The rate at which users like or repost a post, as a function of how high it appeared in their feed when they switched to it (since users may access the feed multiple times with overlap in posts, we use the highest position that the post appeared). There are strong feed positioning effects. We consider only posts, likes, reposts, and accesses up to September 16, 2025, and filter to user likes and reposts that occurred within 30 seconds of the user accessing Paper Skygest. Bluesky typically requests pages of size 10 and 30; we use access data for cases where Bluesky requests 30 posts. This results in 5,201 likes and 1,046 reposts. We take the average across post-user pairs for each rank, and obtain clustered $95\%$ confidence intervals by generating 1,000 bootstrap samples of users. The same pattern holds when we increase the interval considered above 30 seconds (see Appendix \ref{['app:funnel']}).
• Figure 4: The distribution of arXiv categories among (a) posts on Bluesky containing arXiv links, (b) all posts containing arXiv links shown to Paper Skygest users (allowing duplicate posts across users), and (c) posts containing arXiv links liked by Paper Skygest users (again allowing duplicates across users). We show categories that are in the top 10 most common categories for at least one of these distributions. We obtain clustered $95\%$ confidence intervals by generating 1,000 bootstrap samples of users. We limit our consideration to posts from non-bot accounts. We identify bots as accounts which mention arxiv or bot in their handle, or have very high posting activity on Bluesky. We use public data from arXiv arxiv_org_submitters_2024 to find a post's categories given its arXiv identifier. There is heterogeneity in the rate at which posts are interacted with, across categories. Note that bars do not add up to one due to ommitted arXiv categories.
• Figure 5: The rate at which users like or repost a post, as a function of how high it appeared in their feed when they switched to it (since users may access the feed multiple times with overlap in posts, we use the highest position that the post appeared). There are strong feed positioning effects. We consider only posts, likes, reposts, and accesses up to September 16, 2025. Bluesky typically requests pages of size 10 and 30; we use access data for cases where Bluesky requests 30 posts. We take the average across post-user pairs for each rank, and report clustered $95\%$ confidence intervals.