Why We Built SwiftScholar

Looking back, it has been over three years since I first stepped into a university. During that time, I wrote countless course papers, and the number of references in my Zotero library approached 600. At that time, the impact of AI on higher education was just beginning to emerge, and most universities in Hong Kong strictly prohibited the use of AI under any circumstances. I considered myself an old-school person, so I disdainfully chose the most professional, or traditional, method of paper writing. I used Google Scholar to find articles, skimmed the abstract, and downloaded them to Zotero. In Zotero, I could conveniently annotate or copy certain sentences as support for my writing; for articles that required in-depth reference, I even had a habit of printing them out and highlighting with a fluorescent pen. Honestly, looking back now, this approach seems a bit pretentious. The academic workload in Hong Kong is heavy, with almost every course requiring a final paper. That means I had to write five papers each semester, forcing me to engage in intensive reading. Fortunately, the papers were not long, and on average, each paper required selecting 7-8 references from about 20 articles skimmed. So, I needed to read nearly 200 articles a year. I needed a more efficient reading method, especially as a non-native English speaker. Typically, translating English to Chinese is the simplest option. It allows you to read the entire article at native-language speed while preserving all original content, and it's completely free. I think many students did this in the pre-AI era: using Google Translate to translate entire articles into their mother tongue. However, there are some issues. First, full translation often makes many academic terms lose their precise meaning, and you miss the opportunity to become familiar with these terms. Often, after skimming, you still need to go back to the English original to find the corresponding sentences to accurately understand the author's intent; nevertheless, translation does speed up the literature screening process. Another unsatisfactory point is the structure of the original text: even though most academically literate people agree that rigorous writing requires good groundwork and paragraph transitions, honestly, they are not very reader-friendly. I often find myself searching between the lines for key statements or the threads linking data to conclusions. This brings us to AI. When AI first appeared, it was mainly in the form of chat interfaces. Some students would directly throw the paper title to AI, but as someone who considers himself a rigorous scholar, I absolutely disliked doing that. However, I still needed it, so I conservatively had AI summarize the PDF of the paper I provided, and then asked appropriate questions to check if the article was what I wanted. Since AI summarizes the full text, it provides more information than the abstract, especially when the information you want is not just what the author intended to express. We have all learned some paper reading methods, at least "seen the pig run" when preparing for IELTS. Sometimes you need to refer to the research methods, sometimes you need definitions of key concepts, or you need the article to tell you where to find such information. So, in summary, the reading workflow at that time became: read abstract -> discuss details with AI -> read translated paper -> read original text -> annotate for writing. Fortunately, this workflow supported me for two years, helping me achieve a GPA sufficient for further studies in Hong Kong and a scholarship that allowed me to enjoy some time in Shenzhen.

A year after graduation, I returned to the university campus and was amazed to find that AI had made tremendous progress in paper reading. It has moved far away from the previous conversational interaction towards more native application experiences. Some applications continue the style of PDF readers but automatically add highlights and notes (like an automated Zotero); others break down papers into different dimensions and provide brief summaries for cross-paper comparison. But on closer thought, they haven't solved many reading problems; instead, they seem to presumptuously tell you what is important and what is not. If a software automatically marks key points for you, then either you will ignore other possible key points (depending on your research question), or it is meaningless (if you spend time reading the full text, the time saved on highlighting and summarizing is negligible). As scholars, our purpose is to view others' research from our own perspective and ultimately draw reasonable conclusions. Even for an undergraduate completing coursework, this is inevitable. If one blindly relies on AI-generated summaries, it is essentially no different from directly copying from AI—because the student neither learns from reading nor gains further thinking. At this moment, my good friend xy3 introduced me to his demo, which eventually became today's SwiftScholar. I found that this project could concretize the philosophy I have always held: efficiency should not come at the expense of independent thinking. SwiftScholar provides bilingual, plain-text content that adapts to screens of various sizes. For quick skimming needs, it offers mind maps and detailed outlines to quickly grasp the article's structure and decide reading priorities. Moreover, SwiftScholar presents paper content in a novel way: structured reorganization. We aim to "losslessly compress" papers into an understandable, quickly searchable structure while preserving as much original content as possible. The basic parsing template covers common sections such as background, prior knowledge, methods, experimental setup, results, and conclusions. Importantly, the hierarchical content structure plus bullet-point-style paragraphs allow you to locate the information you want at a glance and easily connect the logical flow of the article. The advantage of this system is that, compared to translation, it preserves professional terminology while organizing the structure into a reader-friendly format. Considering the importance of figures and formulas in papers, SwiftScholar has made special adaptations. Currently, any content presented as images in papers is analyzed using multimodal models in context to support the article's arguments. For scholars in science, engineering, and economics, the accuracy of formulas is the bottom line. We not only support LaTeX-level formula rendering but also ensure that formulas do not get disordered in various layouts. This system has now completely replaced my previous reading workflow. In last semester's coursework, I completed reading 15 papers in one day, supporting the theoretical foundation of a 2000-word essay. Of course, as the most in-depth user of SwiftScholar, we are continuously polishing this product. In future plans, we will continue to improve the close reading and reading experience, striving to build a one-stop paper reading workflow.

Welcome everyone to use SwiftScholar.