怎么读一篇AI论文How to Read an AI Paper

先找研究问题：引言里「这篇要解决什么」那一两句。找不到，论文自己就没说清，谨慎往下读。

摘要只当目录不当结论：NeurIPS 官方清单要求作者自查「摘要与引言的主张是否准确反映论文的贡献与范围」——摘要是承诺，证据在正文的表里。

定位主证据：先回答「核心结论依据哪张表/哪张图」，再看那张表的设置——什么基准、什么基线、什么指标，最后才看数字。

直接读局限节：NeurIPS 清单同样要求作者描述局限。作者自己写的局限就是适用范围的边界，比任何第三方解读可靠。

分清「方法新」与「结果好」：有消融实验才能看出哪个组件在起作用；没有消融，提升的来源就是未知。

引用数、会议录用、发布热度都不是质量依据；看到 SOTA 声明先查数据污染，再查是否同口径对比。

Find the research question first: the one or two sentences in the intro saying what the paper sets out to solve. If you can't find it, the paper itself isn't clear — read on cautiously.

Treat the abstract as a table of contents, not a conclusion: NeurIPS's official checklist makes authors self-audit whether 'the main claims in the abstract and introduction accurately reflect the paper's contributions and scope' — an abstract is a promise; the evidence lives in the tables.

Locate the main evidence: first answer 'which table/figure does the headline claim rest on', then check that table's setup — which benchmark, which baseline, which metric — and only then look at the numbers.

Read the limitations section directly: the NeurIPS checklist also asks authors to describe limitations. What authors admit themselves is the boundary of applicability — more reliable than any third-party take.

Separate 'new method' from 'good results': ablations are what show which component actually matters; without ablations, the source of the gain is unknown.

Citation counts, conference acceptance, and launch buzz are not evidence of quality; on seeing a SOTA claim, check benchmark contamination first, then whether the comparison is apples-to-apples.