We turned 10 years of ML papers into one page per topic: benchmark records, what got tried, what reviewers expect [P]

We've been working on this for a while and shipped it today. We took every accepted paper from ICML, NeurIPS, ICLR, CVPR and AAAI since 2016 (about 109k), grouped them into ~1,260 topics, and pulled the results tables and ablations out of the full text of 79k of them. The research map just shows the grouping, but each topic page shows what the papers actually report. Here's the one for compact/efficient Gaussian splatting: https://atlas.scime.ai/research/3d-vision/gaussian-splatting/compact-and-efficient-gaussian-splatting A few things it shows: - the topic went from 29 papers in 2023-24 to 24