XPeng’s FastDriveVLA Achieves 7.5x AI Compute Reduction with AAAI 2026 Acceptance
XPeng and Peking University developed FastDriveVLA, a visual token pruning framework that reduces autonomous driving AI’s computational load by 7.5x while maintaining high planning accuracy on the nuScenes benchmark. The research was accepted by AAAI 2026, which selected 4,167 papers from 23,680 submissions (17.6% acceptance).
1. Arrowpoint Exits XPeng Stake
According to an SEC filing dated Nov. 13, 2025, Arrowpoint Investment Partners (Singapore) liquidated its entire position in XPeng Inc., selling 500,000 shares for an estimated $8.94 million based on quarterly average prices. This holding had represented roughly 8 percent of the fund’s assets under management at the end of the prior quarter, and post-transaction the stake stands at zero. In the wake of the sale, Arrowpoint’s top equity allocations are led by AEG (15.4 percent of AUM), followed by SATS (3.8 percent), VCSH (3.4 percent), ALAB (3.4 percent) and ATAT (2.7 percent), underscoring the fund’s portfolio diversification strategy. While XPeng’s stock had more than doubled year to date by Q3, this full exit may reflect a tactical profit-taking decision rather than a loss of confidence, given the company’s 156 percent year-over-year delivery growth through November and an export volume now accounting for about 10 percent of total shipments.
2. XPENG-PKU FastDriveVLA Research Accepted by AAAI 2026
In collaboration with Peking University, XPeng’s FastDriveVLA paper has been accepted for presentation at AAAI 2026, one of the field’s most selective conferences. Out of 23,680 submissions, only 4,167 papers secured acceptance, reflecting a 17.6 percent rate. FastDriveVLA introduces a reconstruction-based visual token pruning framework that reduces onboard computational load by a factor of 7.5 while preserving high planning accuracy on the nuScenes autonomous driving benchmark. The method trims the visual token count from 3,249 to 812 by emulating human drivers’ focus on critical foreground elements, leveraging an adversarial foreground-background reconstruction strategy. This achievement builds on XPeng’s full-stack AI capabilities and marks its second top-tier conference recognition this year, furthering the company’s push toward scalable Level 4 autonomy deployment.