Tsinghua AutoSOTA spends about $104K in a week to produce 105 SOTA results

刷榜只是体力活！清华消费10万块，一周「肝」出105个SOTA

Tsinghua's Fengli Xu team and Beijing Zhongguancun Academy released AutoSOTA, which ran unattended for one week, used about 22B tokens, and produced 105 SOTA results. The system uses eight agents for resource setup, environment fixes, scheduling, idea generation, and audits; each full run averaged 5 hours. The key check is its red-line audit: it forbids changing evaluation scripts and data splits, which decides reproducibility.

Why it matters: HKR-H/K/R all pass: hard numbers, an 8-agent mechanism, and audit constraints make the claim testable. It stays at 84 because this is single-source secondary coverage, not a major model or product release.