XPeng Inc. (XPEV), working alongside Peking University, announced Monday that its autonomous driving research paper earned acceptance to AAAI 2026, a pretty big deal in artificial intelligence circles. The conference received 23,680 submissions this year and accepted just 4,167 papers, which works out to a highly selective 17.6% acceptance rate.

Teaching Cars to Focus Like Humans

The paper introduces FastDriveVLA, which tackles a fundamental problem in autonomous driving: current AI systems process way too much visual information. The framework uses a visual token pruning method that helps self-driving systems concentrate on critical visual cues while filtering out irrelevant background noise, much like how human drivers naturally focus on what matters.

Here's why this matters. Vision-Language-Action models have become increasingly popular in autonomous driving because they're excellent at understanding scenes and reasoning about actions. The catch? They typically process thousands of visual tokens per image, which creates massive computing demands and slows down the real-time decision-making that's absolutely essential when a car is moving at speed.

Dramatic Efficiency Gains

FastDriveVLA solves this through what the researchers call a reconstruction-based pruning method inspired by actual human driving behavior. The framework uses an adversarial foreground-background reconstruction strategy to identify and keep only the high-value visual tokens.

The results from testing on the nuScenes autonomous driving benchmark are impressive. When visual tokens were reduced from 3,249 down to just 812, the model achieved nearly a 7.5-fold reduction in computational load while maintaining high planning accuracy. That's the kind of efficiency gain that actually matters for putting these systems in real vehicles.

Building Toward Full Autonomy

This represents XPeng's second appearance at a major global AI conference in 2025. The company presented at CVPR WAD back in June, and at its AI Day in November, XPeng unveiled its VLA 2.0 architecture, which streamlines the process by removing the language translation step to enable direct Visual-to-Action generation.

The Tesla Inc. (TSLA) rival says these advances strengthen its full-stack AI capabilities as it pushes toward L4 autonomous driving, the level where vehicles can handle most driving situations without human intervention. XPeng shares have climbed 76% year-to-date, powered by strong demand for its electric vehicles.

XPEV Price Action: XPeng shares were down 2.12% at $20.34 during premarket trading on Monday.