Framework Development

XPENG–Peking University’s FastDriveVLA Accepted at AAAI 2026, Cutting VLA Compute Nearly 7.5x

The reconstruction-based token pruning method reports state-of-the-art nuScenes results with far fewer visual tokens.