humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Models & learning

Scaling law

A machine-learning scaling law is an empirical relationship between training scale, such as dataset size, model size, or compute, and a measured outcome, commonly model loss. It can guide estimates within the tested setup but does not guarantee similar gains for a different metric or deployment setting.

Also known as: machine-learning scaling law, data scaling law, model scaling law

Updated

What the relationship measures

A scaling law summarizes how a chosen model metric changes as one training dimension grows. A study may vary dataset size, model parameters, or training compute. The classic language-model scaling study by Kaplan and colleagues measured how cross-entropy loss changed with model size, dataset size, and compute. Those relationships were measured for particular models and training setups; they are not a guarantee for every system.

Reading a robotics scaling claim

A robot study needs to say which variable changed and which outcome was tracked. Figure's Helix 2.5 announcement says four models were trained on Index subsets spanning an eightfold data increase, with model size and downstream training fixed. The outcome was held-out robot-action prediction loss. Figure reports that the largest run's test loss could be forecast from smaller runs with an error equal to 0.54% of the variation across that tested range.

That result concerns the prediction metric in that setup. It does not show that full-task success improves by the same percentage, or that a curve measured on one model and dataset can be extrapolated to another robot or task. Compare scaling claims with a separate held-out evaluation of the robot policy and the robot generalization conditions that matter in deployment.

Sources