from Hacker News

Ask HN: How to discover new and interesting papers?

by drsim on 9/17/23, 1:32 PM with 10 comments

I’ve recently gotten into reading scholarly papers, prompted by some very accessible and engaging papers about LLMs.

Most of these are hosted at arXiv which has a new function, but no ranking.

I’d love to discover interesting papers in the fields I follow, ranked by the community. Much like HN.

Does this exist? How do others keep up to date in their fields?

  • by evolve2k on 9/17/23, 9:25 PM

    I often use semantic scholar (https://www.semanticscholar.org) as my first place to search on new topics.

    Once you get the search results up you can then refine and sort by: relevancy, recency as well as “citation count” and “most influential papers”.

    The later I find especially useful when exploring a new topic.

    They have a login account but to be honest I haven’t really explored what that offers. Anyone have an account and finding that useful?

  • by strangecasts on 9/17/23, 1:59 PM

    For fields where ArXiv preprints are the norm, Arxiv-Sanity [1] allows filtering by keywords and basic topic filtering.

    In ML, I personally like skimming Davis Blalock's Davis Summarizes Papers [2].

    [1] https://arxiv-sanity-lite.com/

    [2] https://dblalock.substack.com/

  • by austinjp on 9/18/23, 6:14 AM

    Here are a few options to consider. First, Google Scholar. If you're logged into Google it will make a handful of recommendations on its front page. I've not really paid attention to how good the recommendations are. It says they're based on your Google Scholar record and alerts, so I guess you'll need both/one of those for it to work.

    https://scholar.google.com

    Second, Scopus from Elsevier (a company that plenty of people dislike). You'll need to create an account, and I don't know if non-academic accounts have the same access as academic ones. It has a new "researcher discovery" function I've not used so again can't vouch for its quality. You can set up various alerts apparently, although again I've not used them.

    https://scopus.com

    If an author is registered on ORCID you can check their works, but it doesn't appear that anything like RSS feeds are available, unfortunately. Plenty of journals have RSS feeds, but you'll have to hunt them down yourself.

    https://orcid.org

    Finally, you might want to check out other platforms and preprint servers, which might have better alerts etc. Try OSF, which hosts a bunch of preprint servers, and also provides hosting for documents and files that accompany published papers. However, it looks like there isn't much comp-sci stuff on there.

    https://osf.io

    I guess you could have a look at figshare.com too for similar reasons.

  • by vlmutolo on 9/17/23, 7:31 PM

    I’ve used Litmaps[0] to discover new papers in a field. They have an interesting “discover” mode where you input papers you consider to be “relevant” and they try to suggest other papers you’d think are relevant.

    If you subscribe to their service, they’ll even notify you when new papers come out that you’d consider relevant.

    [0]: https://www.litmaps.com/

  • by rewmie on 9/17/23, 2:07 PM

    As for ranks, one of the main metrics is citations. Keywords are also relevant. Once you are up to speed on a topic then you'll notice a few publications tend to publish the good stuff, and so subscribing to them helps you feed off the paper firehouse.

    arXiv might be a source but hosting only preprints without peer review ends up lowering the average quality of the stuff that shows up in there.

  • by jrod16 on 9/18/23, 11:33 AM

    You can use a service like usedigest.com and add in RSS feeds of specific sources related to scholarly papers. Then you’ll get a personalized newsletter with that content every day. I’ve been using it to keep track of top posts on HN and Reddit lately and it’s a huge time saver.

  • by astrocato on 9/18/23, 11:00 AM

    Check out dailyacademia.com. It delivers a daily digest of new publications from arXiv based on your interests. I've set up a promo code "hackernews" for a free one-year subscription.

  • by janalsncm on 9/18/23, 11:15 AM

    Look into Papers With Code. It has a section on state of the art which contains papers on all sorts of ML tasks and the models which are best at that task.

  • by py4 on 9/17/23, 7:07 PM

    aside from sources mentioned by others (arxiv-sanity-lite, newsletter): 1. deep learning monitor: https://deeplearn.org 2. Following folks on Twitter and then Twitter recommendation algo will take care of the rest