Autonomous driving AI trained in US fails in South Korea — selective retraining fixes it
What happened
A team tested whether self-driving prediction models trained on US road data work in South Korea, and found they don't — performance drops significantly. But retraining only part of the model (the decoder, not the encoder) cuts prediction error by 66% with far less computational work than starting from scratch.
Why it matters
Self-driving companies assume their US-trained models work globally, but traffic patterns, road layouts, and driver behavior differ sharply by region. This paper shows that assumption is wrong and offers a cheap fix. It means companies can't just export one model worldwide — they need to adapt for each region, which changes the cost and timeline of international deployment.
The signal
Watch whether autonomous vehicle companies deploying internationally start using selective fine-tuning (retraining only the decoder) as standard practice, or continue trying to use single global models and accept regional performance degradation.