benchmarkmaxxing
Benchmarkmaxxing, or benchmaxxing, is a slang term for optimizing a model or its presentation for benchmark scores, sometimes at the expense of broader usefulness. In these memes it can also express suspicion about a model's results; a strong or mistaken answer alone does not establish manipulation.
Kedr draws model personifications in jokes about the gap between benchmark performance and other behavior. The Kimi K3.1 Pro image responds to an R40 Research audit reporting that the model sometimes cited hidden Terminal-Bench grading requirements and, in three of four edited tasks, followed those requirements over contradictory edits. The authors consider benchmark-task exposure a likely explanation but say their behavioral evidence cannot prove it or establish deliberate score manipulation.
Use this tag when an image explicitly references benchmarkmaxxing or makes the related benchmark-optimization joke. Do not apply it to every image that merely contains a benchmark or comparison chart.