← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

Social Media Feed Elicitation

Lindsay Popowski , Xiyuan Wu , Xuyang Zhu , Tiziano Piccardi , Michael S. Bernstein

DOI PDF Program page Page JSON

This is a strong CHI-style method paper: the novelty is in turning feed customization into an elicited authoring task, and the study design is credible for showing that the interaction helps users produce better-specified feeds. The main caution is that the evidence is still about immediate preference outcomes in a constrained setting, not durable real-world deployment.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

Social Media Feed Elicitation is best read as a method contribution for end-user customization rather than as a new ranking algorithm. Its central move is conceptually simple but practically important: instead of assuming users can directly enumerate all the preferences that should govern a custom feed, the system uses an interview process to surface missing constraints, gaps, and edge cases. That is a meaningful departure from the common-sense view that feed personalization is mainly a matter of asking for the desired content and then executing it. The paper’s novelty lies in the interaction technique and the authoring workflow, not in a new model architecture. The evaluation basis is also reasonably clear: a between-subjects online study with N=400, comparing elicitation against manual description and an interview-less ablation, which is an appropriate design for testing whether the elicitation step improves the resulting feed specification and user preference. At the same time, the evidence is bounded. The study uses a static Bluesky post inventory and immediate judgments, so it does not establish long-term adoption, resilience to changing preferences, or behavior in a live social-media environment. The paper itself acknowledges important limitations: the system has difficulty eliciting and representing prioritization, and the inventory is limited by computational and time costs because retrieval and ranking rely on LLM calls. Those constraints matter because they define the boundary of the contribution: this is strongest as a proof that guided elicitation can improve custom-feed authoring under controlled conditions, and weaker as evidence that the approach fully solves feed control in the wild. Overall, the paper is a solid honorable-mention-level CHI contribution because it combines a clear user problem, a plausible interaction design, and a credible controlled evaluation, while remaining appropriately modest about what the system can and cannot yet do.

What Changed

Canon before

Prior work and common practice assume users can directly specify desired feed behavior or preferences, but this paper argues that users often miss gaps and edge cases in their own feed definitions, motivating an elicitation step before feed generation.

Departure from common sense

The paper challenges the intuitive idea that a user can simply tell a system what they want in a custom feed and get a good result. Instead, it argues that people often miss gaps and edge cases in their own definitions, so an interactive elicitation process is needed to surface those missing preferences.

Actual novelty

The paper introduces feed elicitation interviews, an interactive method that uses LLM-guided questioning to help users identify missing preferences and articulate a structured custom-feed specification. The contribution is not just a new feed generator, but a user-facing elicitation procedure for authoring feeds.

Evidence

The paper’s core claims are supported by an online between-subjects study with N=400. Participants completed a manual feed description process and were randomized into a feed elicitation interview or an interview-less ablation, producing feeds filtered and ranked according to preferences. The discussion also names concrete limitations around prioritization elicitation and inventory constraints.

Limits

Method limits

The evaluation is an online between-subjects study with a static post inventory and immediate preference judgments, so it supports short-term perceived quality and authoring behavior more than long-term use, robustness, or real-world feed outcomes.

Deployment limits

The system depends on LLM calls for retrieval and ranking, which constrains the inventory due to computational and time costs. The current architecture also struggles with preference prioritization and cannot fully support some feed-level objectives.

Boundary conditions

Results are bounded by the static Bluesky post set, the study’s immediate post-authoring setting, and the current post-level ranking architecture. Claims are strongest for custom feed authoring tasks where users can articulate preferences through guided elicitation.

Manual check: Verify that the full-text span used for evidence matches the extracted text and that the limitation wording aligns exactly with the paper’s discussion section.

Position in field

This sits at the intersection of end-user customization, LLM-assisted interaction design, and social media feed control. It reframes feed personalization as an elicitation problem rather than only a ranking or recommendation problem.

Abstract

Social media users have repeatedly advocated for control over the currently opaque operations of feed algorithms. Large language models (LLMs) now offer the promise of custom-defined feeds--but users often fail to foresee the gaps and edge cases in how they define their custom feed. We introduce feed elicitation interviews, an interactive method that guides users through identifying these gaps and articulating their preferences to better author custom social media feeds. We deploy this approach in an online study to create custom Bluesky feeds and find that participants significantly prefer the feeds produced from their elicited preferences to those produced by users manually describing their feeds. Through feed elicitation interviews, we advance users' ability to control their social media experience, empowering them to describe and implement their desired feeds.