Paper Notes

Paper Notes solves a specific, annoying problem: conference season dumps thousands of submissions on you and the abstract tells you almost nothing about whether the method is worth your afternoon. Five-minute notes at least let you sort. The 26,590-note index covers past and upcoming editions alike, so you can scan ICLR2026's 5,307 notes the same way you'd scan ECCV2024's 868. What we actually use it for is directional reading rather than deep reading. The subfield counts are the most useful thing on the site. When NeurIPS2025 shows 143 reinforcement-learning papers alongside 143 on model compression, and ICML2026 shows 116 on compression against 110 on reinforcement learning, that's a comparable signal about where effort is going — one the abstract of any single paper won't give you. Pick it if you're a grad student building a literature-review shortlist, a practitioner tracking whether a topic is heating up across venues, or part of a bilingual lab that wants notes in Chinese or English without a signup. The English domain exists, and the GitHub link means the project is inspirable rather than a black box. Pass if you need figures, formulas, ablations or any peer-grade judgment about a method. The notes are AI-assisted by the site's own admission, and the disclaimer is worth taking literally — numbers and equations can drift. Anything you cite should be checked against the paper. There's no API, no export and no automated pipeline, so it doesn't slot into a literature-review workflow you're trying to script. Where it bites is scale: with thousands of notes per venue, browsing is a manual activity, and the density counts help only if you already know which subfields matter to you. The closest alternative for many readers is Semantic Scholar or arXiv itself — both