Note We extend concept drift theory to MLLMs and introduce a T-distribution-based adapter that improves vision-language pre-training under both gradual long-tailed shifts and sudden OOD shifts.