Another month, another round of interesting research papers ranging from large language modeling to computer vision. One recent focus is on refining Large Language Models (LLMs). For instance, introducing models like Platypus and the Reinforced Self-Training (ReST) method are the latest attempts to improve alignment with human preferences.
I am curious how you select which arXiv papers to summarize. The torrent of new postings is totally overwhelming! Do you have any tricks or hints to share? Certainly it is not random! ;)
That's a tricky question :). I usually skim arxiv releases and bookmark those that sound interesting. I also bookmark some papers here and there if I stumble upon them on social media or things that come up in discussions with colleagues. I usually go back to my bookmark list and try to read or skim a few papers each week. I skim more papers than I include here, and I don't skim or read all the papers I bookmarked. I kind of go by what's generally interesting and/or interesting to me.
Certainly sounds plausible. Plus, your interest monitor is much better than mine. Wonder what your false-negatives (ie, key papers that appear boring) are? Could you monitor your hit-rate by citation rates after a month or year? Are you using Semantic Scholar or Connected Papers? Wonder if you could be replace (or do a better job) with a properly tuned LLM? This flow selection is a BIG Problem. Maybe the solution to maintaining sanity is to not read any paper that is less than six months ago, having then proved its worthiness.
What about... Scan new arXiv postings. Summarize and embed in a latent space customized for arXiv. Look around its neighborhood and rate nearby papers that are semantically similar. The higher rating get our attention. NO... somehow we need to judge that there is some novel idea/twist to previous research.
Hmmm... Must be nearby other important papers but also carve out your own new space. Maybe do a 'diff' with nearby paper embeddings and (somehow) rate the 'novelty' of the new paper. Getting messy! Guess I will rely on your trusty interest monitor ;)
Regarding the date range, I strictly stick to the date range defined in the title. So, older papers wouldn’t make the cut for this recent research series 😅.
I would definitely do different types of literature searches eg via Google Scholar etc when I work on a specific project. But yeah, here it’s strictly a simple “what’s new this month” kind of overview.
I am curious how you select which arXiv papers to summarize. The torrent of new postings is totally overwhelming! Do you have any tricks or hints to share? Certainly it is not random! ;)
That's a tricky question :). I usually skim arxiv releases and bookmark those that sound interesting. I also bookmark some papers here and there if I stumble upon them on social media or things that come up in discussions with colleagues. I usually go back to my bookmark list and try to read or skim a few papers each week. I skim more papers than I include here, and I don't skim or read all the papers I bookmarked. I kind of go by what's generally interesting and/or interesting to me.
Certainly sounds plausible. Plus, your interest monitor is much better than mine. Wonder what your false-negatives (ie, key papers that appear boring) are? Could you monitor your hit-rate by citation rates after a month or year? Are you using Semantic Scholar or Connected Papers? Wonder if you could be replace (or do a better job) with a properly tuned LLM? This flow selection is a BIG Problem. Maybe the solution to maintaining sanity is to not read any paper that is less than six months ago, having then proved its worthiness.
What about... Scan new arXiv postings. Summarize and embed in a latent space customized for arXiv. Look around its neighborhood and rate nearby papers that are semantically similar. The higher rating get our attention. NO... somehow we need to judge that there is some novel idea/twist to previous research.
Hmmm... Must be nearby other important papers but also carve out your own new space. Maybe do a 'diff' with nearby paper embeddings and (somehow) rate the 'novelty' of the new paper. Getting messy! Guess I will rely on your trusty interest monitor ;)
Regarding the date range, I strictly stick to the date range defined in the title. So, older papers wouldn’t make the cut for this recent research series 😅.
I would definitely do different types of literature searches eg via Google Scholar etc when I work on a specific project. But yeah, here it’s strictly a simple “what’s new this month” kind of overview.