How to Build a Personalized Research Feed That Learns from You

Information is not scarce. That stopped being the problem a long time ago. ArXiv publishes thousands of new preprints every week. PubMed indexes over a million articles a year. Semantic Scholar covers hundreds of millions of papers across every discipline. The raw material for your next breakthrough is out there, sitting in a database, waiting to be found.

The real problem is that almost none of it is relevant to you. And the small fraction that is relevant — the handful of papers that would genuinely change how you think about your research — is buried under an avalanche of tangentially related work. What you actually need is not more access to papers. You need a personalized research feed that is built around your specific interests and, more importantly, keeps evolving as those interests change.

This article explains how to build exactly that, step by step.

The Problem with "Set It and Forget It" Tools

Most researchers have tried some version of the same approach. You set up Google Scholar alerts with a few keywords. Maybe you subscribe to a handful of RSS feeds or follow some Twitter accounts. You bookmark a couple of "awesome lists" on GitHub. For the first week, it feels productive. Then life happens.

The fundamental flaw with traditional tools is that they are static. You configure them once, and they never change. But your research interests are not static. A PhD student who started working on attention mechanisms in 2023 might be deep into state-space models by 2025. A bioinformatician who began with single-cell RNA-seq might now be exploring spatial transcriptomics. Research is a moving target, and your discovery tools need to move with you.

Here is what typically goes wrong:

Keyword alerts become stale. You set them up during one phase of your research, and they keep delivering papers for a version of your interests that no longer exists.
Manual curation does not scale. Following specific journals, authors, or conference proceedings requires constant attention. When you get busy — and you will get busy — the whole system collapses.
There is no feedback loop. The tools you use have no way of knowing which papers you found useful and which ones wasted your time. Every day, they give you the same quality of results regardless of your engagement.

What you need is a personalized research feed that does two things: starts with what you care about right now, and continuously learns what you actually find valuable.

Two Layers of Personalization That Actually Work

The most effective approach to building a personalized research feed is not about finding one perfect algorithm or one perfect source. It is about layering two distinct types of personalization on top of each other so that each one compensates for the weaknesses of the other.

ZiNote is a mobile app built around exactly this principle. It uses a Tinder-style swiping interface for paper discovery, but the real power is in how its two-layer personalization system works under the hood.

Layer 1: Keyword-Driven Search Across All Sources

The first layer is straightforward but important. You enter the keywords that define your current research interests, and the system automatically searches across every major academic source — arXiv, PubMed, Semantic Scholar, and others. You do not need to pick which databases to search, configure API connections, or remember which journal publishes what. The system handles all of that.

This matters more than it might seem. One of the hidden time sinks in academic paper discovery is source fragmentation. A relevant paper might appear on arXiv as a preprint, on PubMed as a published article, and on Semantic Scholar with citation context. Most researchers default to searching one or two sources and miss what is available elsewhere. A good personalized research feed removes that friction entirely.

The keyword layer gives you a strong starting point. But keywords alone have a ceiling. They capture what you can articulate about your interests, not what you actually respond to when you see it. That is where the second layer comes in.

Layer 2: Behavior Learning Through Every Swipe

This is where the system becomes genuinely personal. Every time you interact with a paper — every swipe, every save, every skip — you are generating a feedback signal. The system uses these signals to continuously refine what it shows you next.

Here is how it works in practice:

You right-swipe three papers about graph neural networks. The system registers that this topic is resonating with you right now and begins boosting papers in that direction. You will see more work on GNNs, related architectures, and applications in your subsequent feed.
You left-swipe five papers about NLP survey articles. The system picks up on the pattern and automatically reduces the frequency of broad survey papers in your feed. No manual adjustment required.
Over time, the feed converges on your actual preferences. Not the preferences you think you have, or the ones you described with keywords three months ago, but the ones revealed by your real behavior, right now.

This is the critical difference between a static tool and a true personalized research feed. The static tool asks you what you want once. The adaptive system watches what you do and draws its own conclusions, continuously, without you needing to touch a settings page.

The beauty of combining these two layers is that they handle different failure modes. Keywords are great for cold starts and for explicit, well-defined interests. Behavior learning is great for discovering adjacent topics you did not know you cared about and for adapting when your interests shift faster than you update your keyword list.

How to Cold-Start Your Feed in Three Days

Any recommendation system faces the cold-start problem: it needs data about you before it can personalize effectively, but it cannot get that data until you start using it. Here is a practical approach to getting through that phase as quickly as possible.

Day 0: Enter 3 to 5 Precise Keywords

The quality of your initial keywords matters more than the quantity. The goal is to be specific enough that the system can find papers in your actual niche, not the entire field around it.

Too broad: "artificial intelligence," "machine learning," "biology"

Just right: "diffusion models for protein structure prediction," "graph neural networks for drug discovery," "transformer architectures for time series forecasting"

Think of each keyword as a search query you would type if you were looking for your ideal next paper to read. If a keyword would return ten thousand results on Google Scholar, it is too broad. If it would return fifty to five hundred, you are in the right range.

Three to five keywords is enough to generate a diverse but relevant initial feed. You can always add more later, and the system will suggest new keywords when your current ones begin to exhaust their paper supply.

Day 1: Swipe Through 20 to 30 Papers

This is the most important step in the entire process. Your first session of active engagement gives the system the baseline data it needs to start learning. Spend fifteen to twenty minutes swiping through papers. Do not overthink it — your gut reaction to a title and abstract is exactly the signal the system needs.

Right-swipe the papers that genuinely interest you. Left-swipe the ones that do not. The system is not looking for perfection; it is looking for patterns.

Day 3: Notice the Difference

By your third day of regular use, the feed should feel noticeably different from what you saw on day one. Papers that match your actual interests — not just your keywords — will start appearing more frequently. Topics you consistently skip will fade into the background. The feed starts to feel less like a search result page and more like a curated reading list assembled by someone who knows your taste.

This is the inflection point where a personalized research feed stops being a tool you use and starts becoming a tool that works for you.

Completing the Loop: From Discovery to Library

Finding relevant papers is only half the workflow. The other half is doing something useful with them — saving, organizing, annotating, and eventually citing them in your own work.

This is where many discovery tools fall short. They help you find papers but then leave you to manually export, download, and file them into whatever reference manager you use. That friction adds up. Every extra step between "this looks interesting" and "this is saved in my library" is a step where papers get lost.

ZiNote addresses this by offering direct sync with Zotero, the most widely used open-source reference manager in academia. When you save a paper in ZiNote, it can automatically appear in your Zotero library, complete with metadata. No copy-pasting DOIs, no manual PDF downloads, no drag-and-drop from your browser.

The result is a closed loop:

Discover papers through your personalized feed
Evaluate them with a quick swipe
Save the ones that matter
Sync saved papers directly to your reference manager
The system learns from every one of these actions and improves your next session

This loop means your daily paper discovery habit takes minutes, not hours, and nothing slips through the cracks.

Why Your Feed Should Be Smarter Than Your Keywords

There is a deeper principle at work here that is worth making explicit. Keywords represent your conscious research interests — the topics you can name and articulate. But some of the most important discoveries happen at the edges of what you can articulate. You might not have a keyword for "applications of contrastive learning to medical imaging," but when you see a paper about it, something clicks.

A well-designed personalized research feed captures those moments. When you right-swipe a paper that does not match any of your explicit keywords, the system registers that signal and uses it to expand your feed in directions you had not anticipated. Over weeks and months, this means your feed is not just keeping up with your stated interests — it is actively helping you discover new ones.

This is what separates a recommendation system from a search engine. A search engine gives you what you ask for. A recommendation system, when built well, gives you what you did not know you needed.

Getting Started

If you are spending more than thirty minutes a day trying to stay current with the literature, or if you have given up trying altogether, it is worth rethinking your approach from the ground up.

ZiNote is available as a free mobile app. The setup takes about two minutes: enter your keywords, start swiping, and let the system learn. Within a few days, you will have a personalized research feed that reflects not just what you study, but how you think.

Your research interests are too important — and too dynamic — to be served by a static list of alerts. Build a feed that evolves with you.

Download ZiNote and start building your personalized feed today.