📊 Bigger is not necessarily better.
Across ProteinGym benchmarks, scaling protein language models does not consistently improve zero-shot fitness prediction.
We discuss why model likelihood is not the same as fitness, and what may be limiting performance.
