State of the Art (SOTA)

The best publicly reported result on a given task or benchmark, commonly abbreviated SOTA.

“State of the art” originally refers to the highest level of technical achievement a field has reached at a given time; the English phrase already appears in engineering literature from the early 20th century. In machine learning papers, SOTA refers specifically to the best current result on a public benchmark under the same evaluation protocol — for example, “achieves SOTA on LIBERO.” It's a convenient yardstick for comparing methods, but three caveats apply: it only holds for a specific benchmark and setting, and may not carry over to different data or a different robot; embodied AI's simulated benchmarks saturate easily, so a lead of a fraction of a percentage point can be within statistical noise; and real-robot results are hard to compare across labs because of differences in setup and objects. The website Papers with Code used to aggregate leaderboards for many tasks; that domain now redirects to Hugging Face's papers page.

ExampleIn 2025, OpenVLA-OFT reached 97.1% average success rate across LIBERO's four task suites, which the paper describes as a new SOTA on that benchmark.

Also called

SOTA, SotA, state-of-the-art

Related

Benchmark · Benchmark Saturation · Leaderboard Chasing · Success Rate · Ablation Study

Sources

Wikipedia: State of the art
Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success (OpenVLA-OFT)
paperswithcode-data issue #116: paperswithcode.com now redirects to huggingface (2025-08)

As of

2026-09