English (unofficial) translations of posts at kexue.fm
Source

I Developed an Auxiliary Website for Browsing Papers: Cool Papers

Translated by Gemini Flash 3.0 Preview. Translations can be inaccurate, please refer to the original post for important stuff.

Jianlin Su December 25, 2023 #9907

Introduction

For a long time, I have had the habit of browsing arXiv daily to keep up with the latest achievements in the field as much as possible, reminding myself that "not to advance is to go back." Previously, many readers have asked me how I browse arXiv and what auxiliary tools I use. In fact, for a long time, I directly browsed the official arXiv website without using any algorithmic filtering, going through papers one by one myself. This process is tedious but not unacceptable. The reason I did not use algorithms for initial screening was mainly due to the concern of missed recalls; after all, the purpose of "browsing" is to stay updated, and once an algorithm misses a paper, one "loses the initiative."

Since the release of Kimi Chat, I have been planning to create an auxiliary website combined with Kimi to accelerate the process of browsing papers. Having been slightly more free in the past few weeks, I have initially completed this website with the help of GPT-4 and Kimi. After several days of testing and optimization, it has gradually become stable, so I officially invite readers to try it out.

Cool Papers: https://papers.cool

The Core Details

As the name "Cool Papers" suggests, this website aims to turn browsing papers into a very "cool" immersive experience. Of course, the current implementation is still quite simple; the "coolness" is mainly reflected in the FAQ section for papers answered by Kimi, which allows us to understand the main content of a paper more accurately and efficiently (compared to just looking at the title and abstract), thereby judging whether it is a paper that needs intensive reading.

It is particularly important to point out:

  1. This is a website for "browsing" papers, not for "reading" them. "Browsing" means "screening" and "filtering." The purpose of "browsing" is to find papers that require intensive reading, not to replace intensive reading itself.

  2. Currently, it only supports arXiv as a paper source, synchronizing the latest day’s paper list from arXiv. Therefore, browsing Cool Papers is essentially equivalent to browsing arXiv. Other paper sources, such as OpenReview, may be integrated in the future depending on subsequent usage and feedback.

  3. Because it is positioned for "browsing," the motto is "persistence is key, and time waits for no one." Therefore, it currently only supports displaying the latest day’s papers and does not support historical backtracking for now. Of course, this can also be modified based on future reader feedback and needs.

  4. The FAQ is based on Kimi Chat; please be grateful and cherish it. In fact, many papers can be filtered out based on the title and abstract alone. The existence of the Kimi FAQ is to make more accurate judgments on uncertain papers, so please do not click [Kimi] indiscriminately.

  5. Clicking [PDF] allows you to preview the paper content (limited to PC browsers; mobile browsers will trigger a download). However, this depends on your own network to access arXiv, so if the PDF preview does not appear for a long time, it may be due to your own network issues.

  6. Clicking [Copy] will copy the basic information of the paper (title, abstract, link, etc.) to the clipboard, which can be pasted elsewhere to share the paper.

  7. The paper list maintains the arXiv publishing order by default. If -sorted-by-stars is added, it will be sorted according to the stars calculated based on the clicks of all users.

  8. Paper Updates: Paper updates are directly synchronized with the official arXiv website, with a normal delay of no more than 10 minutes. arXiv updates usually occur around 10:00 AM (Beijing Time) on weekdays, but there may be fluctuations of several hours. Furthermore, there are no updates on Saturdays, Sundays, and some US holidays. If you find that the papers on the website have not been updated, you can check the update status on the official arXiv website.

Further explanation regarding the Kimi FAQ:

  1. After clicking the [Kimi] button corresponding to each paper, you will enter a queue.

  2. If there are many people in the queue, a result like Pending:xxx will be displayed, where xxx is the queue number. This number will update automatically.

  3. Once the queuing is complete, Loading:xxx% will be displayed, followed by the streaming output of the FAQ content.

  4. You do not need to keep the page open during the queuing and generation process. Even if the page is closed, the queuing and generation will continue in the background. Reopening the page and clicking the same [Kimi] button will restore the original progress.

  5. If you click [Kimi] and find that the FAQ content is output instantly, it means the paper has already been read by another reader and the FAQ has been cached.

  6. Because there is a queuing mechanism, clicking [Kimi] indiscriminately will not put too much pressure on the website, but it will make others wait too long, which is unfriendly behavior.

Conclusion

Finally, I welcome everyone’s opinions and suggestions, and you are also welcome to continue proposing requirements. Obviously, the current Cool Papers is still very rough and far from being as "cool" as imagined. Using Kimi for FAQs is only a very basic way of combining Kimi with papers. Kimi’s ultra-long context should still have vast room for imagination. Therefore, I look forward to everyone using their imagination to find a more perfect way to combine Kimi with browsing papers.