Baseline

An existing or simple method used as a point of comparison to show how much a new method improves.

A baseline is the reference point used when evaluating a new method. Google's machine learning glossary defines it as a reference model used to compare against a (usually more complex) model. A baseline can be something simple, like plain behavior cloning or a random policy, or it can be a strong, widely recognized method, like OpenVLA. Reporting “80% success rate” on its own says little; the improvement only means something once it's measured against a baseline under the same benchmark and evaluation protocol. When reading a paper, check whether the baseline is strong enough and whether it was reproduced faithfully to the original authors' settings — a weak baseline makes an improvement look bigger than it is. An ablation study can also be seen as comparing a method against a version of itself with one component removed, which then serves as the baseline.

ExampleThe OpenVLA-OFT paper uses the original OpenVLA as its baseline, raising average success rate on the four LIBERO task suites from 76.5% to 97.1%.

Related

Ablation Study · Benchmark · State of the Art (SOTA) · Success Rate · Evaluation Protocol

Sources

Google Machine Learning Glossary: baseline
Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success (OpenVLA-OFT, arXiv 2502.19645)