CrossWeave: Bridging Perspectives Across Online Communities with a Dual-Pane Design

Social media has been blamed for destabilizing democratic societies. Large majorities in several countries, and especially in the U.S., believe that social media has made us more divided in our political opinions, less civil in our political discourse, and more vulnerable to misinformation [7]. Yet social media is profitable, popular, and likely here to stay. So we must meet users and companies where they are. The challenge is to mitigate problems such as echo chambers [8, 24], unchallenged misinformation [13, 27], selective exposure [3, 23], and perceived polarization [14]—and in an appealing way.

We introduce CrossWeave, a design that aims to broaden the information diet of social media users, connecting them to a wider range of high-quality ideas that bear on whatever they are currently reading or writing. It does not intervene in the main social media feed, but rather augments it with a curated side pane of social media posts from across various platforms. Construction of such a side pane is only now becoming possible, using large language models (LLMs) to automatically identify diverse relevant content and highlight connections. We outline the AI challenges in section 5, but this paper focuses on the UI concept and its prosocial goals. A demo video is at https://youtu.be/QFW9i7KoBhs.

Two screenshots of the CrossWeave interface, showing “Reading Mode” and “Writing Mode” for a focused post arguing that the U.S. is trapped in an "Innovator's Dilemma" with oil profits versus renewable energy. (a) Reading Mode: In the main pane, a sentence about the U.S. being unable to let go of "fat oil profits" is highlighted. In the side pane, a related post by Ed Jones about Wall Street's focus on "short term economic efficiency" is shown with a matching highlight and a tooltip reading "7 more Short Termism," indicating that other side-pane posts address the same theme. (b) Writing Mode: The user is composing a reply in the main pane (“It's not just about ‘profits’; those who work in oil are gonna lose their livelihood”). Below the draft, an AI-generated “Feedback” box acknowledges the point about livelihoods and suggests mentioning solutions for affected workers. Below that, three simulated responses are displayed, e.g., “It's a good point; we need to ensure those workers get support and opportunities during the energy transition.” Meanwhile, the side pane shows related posts color-coded by contribution type, including tags for "Evidence (Public)," "Connections," and "Framing."

Since the early days of Usenet and Facebook, social media and news comment sites have organized discussions into disjoint threads with different readerships. Arguably, this is the core reason for online echo chambers [8, 24] and for the limited range and quality of many online discussions [26].

CrossWeave is designed to augment rather than replace this popular traditional social media design. Sustained conversations within existing social circles are clearly engaging and presumably serve real needs (see [30]). Users come to social media for many reasons—to maintain offline social connections [29], to follow and react to news [21], to monitor others’ reactions [29], to share their own views and seek validation [18], to attract attention and gain social status [25], to cultivate followers and demolish opponents [15], and so on. They may not be actively seeking information, deep understanding, or new conversation partners [18, 25]. However, by augmenting their display with a side pane (fig. 1), we hope to make them at least peripherally aware of other views and communities and draw them into exploration and engagement [16, 17].¹

CrossWeave is intended as a conduit rather than a referee. All of the content is still written by actual human users: there is no AI voice summarizing comments [28] or “telling me what to think.” The side pane showcases a diverse set of views and supports easy exploration, so we hope users will regard it as interesting and empowering rather than biased or coercive. The side pane is not personalized to a particular user, but only customized to the current context.

CrossWeave aims at both informational broadening and social bridging. First, it can broaden users’ understanding of the current topic by showing them additional perspectives from all angles. Second, it offers them bridges to new social communities [19]: a user can click on any post in the side pane to jump into its original thread and reply if they choose.

Finally, our design strengthens the incentive to write useful posts—the sort that will be promoted via the side pane. A user may put in the effort to rebut a friend's misinformation, knowing that their effort will not be trapped in its original thread, but may be shown again next to similar misinformation when it appears in other threads across the platform (cf. X Community Notes [31]).²

A CrossWeave user begins as in other social media interfaces, by scrolling through a list of top-level posts in the main pane.³ They may click on any post to open the conversation that it is part of. The main pane then changes to show the post with its thread ancestors above it and its replies below it (as in X, Mastodon, or Bluesky). The user may then post their own reply or continue navigating.

A permanent side pane (fig. 1) displays a ranked list of posts whose contributions relate to the currently focused post—that is, the main-pane post that was most recently clicked on or scrolled to. The user can click on any side post to open its thread in the main pane (thus bridging to that related thread).

Our demonstration focuses on conversations about U.S. news (rather than about books or movies, say). Our demo system is populated with reader comments harvested from Reddit, Fox News, and The New York Times.

We invite users of the demo to consider whether CrossWeave could improve civic discourse on social media:

1. Does the system feel trustworthy and usable?
   
2. When you are reading a post, does the side pane add value by genuinely broadening your perspective?
   
3. Would the side pane and the posting feedback help you take more informed and defensible positions on contested political issues, given your own values?
   
4. Would the system help you understand your political opponents’ positions, arguments, and values well enough that you can have more constructive dialogue with them? Would it draw you into such dialogue?
   

Our demo uses precomputed side pane data for each possible focused post. We first use an LLM (currently GPT-5.4) to describe and classify the contributions of each post, to identify a highlightable span for each contribution, and to assess the relevance and added value. Then, from these judgments, a cheaper secondary pipeline (GPT-5.4 and GPT-4.1-mini) induces a small set of subtopic labels to enable the depth-first drilldown. To finalize the ranking, we first use a deterministic, interpretable function to combine relevance and added value into a single score, and then rank greedily via maximal marginal relevance [6], penalizing redundancy in subtopic and contribution type.

While this implementation illustrates the UI vision and confirms that AI can indeed be used to identify and highlight relevant content, it is quadratically expensive and runs offline. Handling live traffic would require CrossWeave to construct each side pane efficiently on demand (at least when no recent version is cached). We have been working on this problem for some time, in parallel to developing the UI design. Challenges include scalability and latency. A reasonable workflow would identify high-quality reusable posts as they arrive, preprocess them by extracting their substantive key points (see the contribution types in section 3.2), semantically index these key points for rapid retrieval from a precomputed topic graph or vector database, and later rerank retrieved posts on the fly using a fast scoring function that considers diversity [6].

Of course, rapidly assembling a diverse list of relevant documents is the central task of information retrieval. IR and NLP researchers have recently been incorporating LLMs into the indexing process [e.g., 10]. Still, our criteria in section 3.1 may be particularly hard to index for.⁹ They seem to require closer reading, deeper reasoning, and more structured human evaluation than traditional IR settings. We find that it is challenging even for a human to assess numerically how relevant a post is, how valuable its contributions are, and how distinct they are from the contributions of other posts. Furthermore, such assessments are value-laden, subjective, and likely gameable. Yet we must take care to make them fairly, so as not to inadvertently suppress some viewpoints or lose user trust on suspicion of doing so. In short, we regard the retrieval of broadening content as a challenging new task for NLP.

Social media is a consequential user interface. CrossWeave is our effort to reimagine it with AI, not by adding a chat pane but by designing new helpful and prosocial features that are quietly enabled by LLMs. CrossWeave attempts to connect isolated social media threads to the broader landscape of viewpoints, arguments, and users, in the hope of giving users a more mature understanding of this landscape and enabling them to participate more constructively through their own posts.