AI models adapt faster to new data without retraining

Researchers found that neural networks can adjust how they process information at test time by tweaking activation functions instead of retraining weights. This lets models stay accurate when they encounter new types of data or corruptions they've never seen before, without needing access to the original training data.

If this scales beyond image classification, it could reduce the computational cost and time required to deploy AI models in real-world conditions where data shifts unpredictably — a fundamental problem that currently forces expensive retraining cycles.