A benchmark result or SOTA claim that looks too good to be true — raising immediate suspicion about cherry-picked evals, contaminated test sets, unreported training tricks, or outright overfitting to the leaderboard. Sus sota is the AI research community's built-in skepticism filter: when a paper claims impossible gains with unusual methodology, limited reproducibility, or results that somehow don't transfer to real-world tasks, the whole thing gets flagged. A healthy reflex in a field where benchmark gaming is a genuine problem.
Ninety-two percent on that benchmark with half the parameters? Sus sota — someone needs to check whether the test set leaked.
No comments yet — say something.
Add your own interpretation of "sus sota".
Viral internet speak — memes, ratios, main-character moments, and the algospeak of every platform from Twitter to Reddit to TikTok comment sections.
See all Internet & Memes slang on Slangora.