ConceptOut-of-Distribution Benchmark / OOD Benchmark
Out-of-Distribution Benchmark
An evaluation methodology that requires a model to generalize to scenarios and interaction combinations not seen in the training distribution. Such benchmarks are designed to expose a model's genuine capabilities rather than its memorization of statistical regularities present in training data, thereby providing a more demanding test of transfer and reasoning.
Timeline (last 90 days)
Oct 12
Evaluating within the training distribution easily overestimates model capability, while out-of-distribution benchmarks require the model to generalize to unseen scene and interaction combinations, better exposing the gap between 'looking right' and 'being physically right'.
Unverified50%