XPENG-Peking University Collaborative Research Accepted by AAAI 2026: Introducing a Novel Visual Token Pruning Framework for Autonomous Driving

XPENG (NYSE:XPEV) and Peking University announced that their paper "FastDriveVLA: Efficient End-to-End Driving via Plug-and-Play Reconstruction-based Token Pruning" was accepted to AAAI 2026.

Rhea-AI Summary

XPENG (NYSE:XPEV) and Peking University announced that their paper "FastDriveVLA: Efficient End-to-End Driving via Plug-and-Play Reconstruction-based Token Pruning" was accepted to AAAI 2026. The research introduces FastDriveVLA, a reconstruction-based visual token pruning framework for Vision-Language-Action models that aims to focus on essential visual information and ignore irrelevant background.

Key metrics: AAAI received 23,680 submissions with a 17.6% acceptance rate; FastDriveVLA cut visual tokens from 3,249 to 812 and achieved a ~7.5x reduction in computational load on the nuScenes benchmark while preserving planning accuracy.

Positive

• ~7.5x reduction in computational load on nuScenes benchmark
• Visual tokens reduced from 3,249 to 812, enabling lighter onboard inference
• Paper accepted at AAAI 2026 with a 17.6% acceptance rate

Negative

• None.

Argus Dec 29 session

-1.35% close to close Open Argus

Details

News Market Reaction – XPEV

In the Dec 29 session, XPEV declined 1.35%, reflecting a mild negative market reaction.

Data tracked by StockTitan Argus on the day of publication.

Market Context

This announcement showcases XPENG’s AI depth, with its FastDriveVLA framework reportedly delivering ...

Analysis

This announcement showcases XPENG’s AI depth, with its FastDriveVLA framework reportedly delivering a 7.5x reduction in computational load while maintaining planning accuracy, and acceptance at the selective AAAI 2026 conference. It builds on prior AI Day and VLA 2.0 disclosures, reinforcing a full-stack autonomy strategy. Investors may track how quickly such research is deployed into production vehicles and whether it supports future margin or performance improvements.

Key Figures

Computational reduction: 7.5x reduction AAAI 2026 acceptance rate: 17.6% AAAI 2026 submissions: 23,680 submissions +4 more

Computational reduction

7.5x reduction

FastDriveVLA visual token pruning framework

AAAI 2026 acceptance rate

17.6%

AAAI 2026 conference paper acceptance rate

AAAI 2026 submissions

23,680 submissions

Total submissions to AAAI 2026

AAAI 2026 acceptances

4,167 papers

Papers accepted at AAAI 2026

Token count before pruning

3,249 tokens

Visual tokens before FastDriveVLA pruning

Token count after pruning

812 tokens

Visual tokens after FastDriveVLA pruning

AAAI recognition count

Second time this year

XPENG recognized at top-tier global AI conferences in 2025

Historical Context

5 past events · Latest: Dec 15

5 events

1. Dec 15

   Global expansion

   24h Move

   -3.6%

   Malaysia localized EV production and strong overseas growth metrics.

2. Dec 01

   Delivery update

   24h Move

   -2.2%

   Robust November deliveries and high overseas growth plus AI Day recap.

3. Nov 17

   Earnings release

   24h Move

   -10.3%

   Q3 2025 revenue, margin improvements, and strong delivery growth.

4. Nov 05

   Earnings preview

   24h Move

   -4.0%

   Scheduling of Q3 2025 earnings call and access details for investors.

5. Oct 31

   Delivery update

   24h Move

   +0.5%

   Record October deliveries and expansion into seven new markets.

24h Move is the share-price change in the day after each event; other market factors may also have contributed.

Key Terms

visual token pruning, vision-language-action (vla) models, end-to-end autonomous driving, adversarial foreground-background reconstruction, +1 more

5 terms

visual token pruning technical

"FastDriveVLA—a novel visual token pruning framework that enables autonomous driving"

Visual token pruning is a technique that trims away small, low-value pieces of an image or intermediate visual features before or during processing by machine-learning models, keeping only the parts that matter for a task. Like pruning dead branches from a tree, it speeds up work and cuts computing costs while aiming to keep the same output quality. For investors, it can lower infrastructure expense, enable faster product responses, and make image‑based services cheaper and more scalable.

vision-language-action (vla) models technical

"framework specifically designed for end-to-end autonomous driving Vision-Language-Action (VLA) models"

Vision-language-action (VLA) models are artificial intelligence systems that can look at images or videos, read and understand text, and take predefined actions based on that combined understanding. For investors, they matter because they can automate tasks like extracting data from charts, reading regulatory filings or product photos, and triggering alerts or workflows—similar to an assistant who can see a chart, read the notes, and press the right buttons—speeding insight while introducing new operational and oversight risks.

end-to-end autonomous driving technical

"VLA models are being widely adopted in end-to-end autonomous driving systems"

end-to-end autonomous driving is a vehicle capability in which the car senses the environment, makes all driving decisions and controls movement without human intervention across an entire trip, not just in specific tasks. For investors it signals potential shifts in costs, revenue models and regulatory risk—like replacing a human driver with a self-sufficient worker—affecting safety outcomes, insurance, fleet economics and how quickly new services can scale.

adversarial foreground-background reconstruction technical

"introduces an adversarial foreground-background reconstruction strategy that enhances the model's ability"

An adversarial foreground-background reconstruction is a deliberate attempt to trick an automated system that separates a main subject (foreground) from its surroundings (background) so the system produces a misleading or altered output. Investors should care because many financial, regulatory and due-diligence tools rely on such automated image, video or document analyses; if those systems can be fooled, it can distort risk assessments, compliance checks and the accuracy of data-driven decisions—similar to someone subtly editing a photo so the important part is hidden or changed.

nuscenes autonomous driving benchmark technical

"On the nuScenes autonomous driving benchmark, FastDriveVLA achieved state-of-the-art performance"

A standardized public dataset and scoring system used to test and compare how well self‑driving software perceives and predicts the road environment. It bundles synchronized sensor recordings (cameras, radar and 3D distance sensors), labeled objects and clear performance metrics so developers get a like‑for‑like ‘exam’ of their perception and planning models. Investors watch benchmark results as a practical signal of technological maturity, safety progress and commercial readiness.

AI-generated analysis. How Rhea-AI works. Not financial advice.

See more from StockTitan in Google Search and AI answers. Adds StockTitan as a preferred source · opens Google

Add on Google Not now

• XPENG-PKU Research Breakthrough: XPENG, in collaboration with Peking University, has developed FastDriveVLA—a novel visual token pruning framework that enables autonomous driving AI to "drive like a human" by focusing only on essential information, achieving a 7.5x reduction in computational load.
• Top-Tier AI Recognition: The research has been accepted by AAAI 2026, one of the world's premier AI conferences, which had a highly selective acceptance rate of just 17.6% this year.
• Accelerating L4 Autonomy: This achievement underscores XPENG's full-stack capabilities in AI-driven mobility and advances the industry toward efficient, scalable deployment of next-generation autonomous driving systems.

GUANGZHOU, China, Dec. 28, 2025 /PRNewswire/ -- XPENG, in collaboration with Peking University, has had its paper "FastDriveVLA: Efficient End-to-End Driving via Plug-and-Play Reconstruction-based Token Pruning" accepted by AAAI 2026, one of the world's top conferences in artificial intelligence. AAAI 2026 received 23,680 submissions, with only 4,167 papers accepted, an acceptance rate of just 17.6%.

The paper introduces FastDriveVLA, an efficient visual token pruning framework specifically designed for end-to-end autonomous driving Vision-Language-Action (VLA) models. This work offers a new approach to visual token pruning by enabling AI to "drive like a human", focusing only on essential visual information while filtering out irrelevant data.

As AI large models evolve rapidly, VLA models are being widely adopted in end-to-end autonomous driving systems due to their strong capabilities in complex scene understanding and action reasoning. These models encode images into large numbers of visual tokens, which serve as the foundation for the model to "see" the world and make driving decisions. However, processing large numbers of tokens increases computational load onboard the vehicle, impacting inference speed and real-time performance.

While visual token pruning has been recognized as a viable method to accelerate VLA inference, existing approaches, whether based on text-visual attention or token similarity, have shown limitations in driving scenarios. To address this, XPENG and PKU developed FastDriveVLA, a novel reconstruction-based token pruning framework inspired by how human drivers focus on relevant foreground information while ignoring non-critical background areas.

The method introduces an adversarial foreground-background reconstruction strategy that enhances the model's ability to identify and retain valuable tokens. On the nuScenes autonomous driving benchmark, FastDriveVLA achieved state-of-the-art performance across various pruning ratios. When the number of visual tokens was reduced from 3,249 to 812, the framework achieved a nearly 7.5x reduction in computational load while maintaining high planning accuracy.

This is the second time this year that XPENG has been recognized at top-tier global AI conference. In June, XPENG was the only Chinese automaker invited to speak at CVPR WAD, where it shared advances in autonomous driving foundation models. At its AI Day in November, XPENG unveiled VLA 2.0 architecture, which removes the "language translation" step and enables direct Visual-to-Action generation, a breakthrough that redefines the conventional V-L-A pipeline.

These accomplishments reflect XPENG's full-stack in-house capabilities, from model architecture design and training to distillation and vehicle deployment. Looking ahead, XPENG remains committed to achieving L4 level autonomous driving to accelerate the integration of physical AI systems into vehicles, with the goal of delivering safe, efficient, and comfortable intelligent driving experiences to users around the world.

About XPENG
XPENG is committed to leading the transformation of future mobility through technological exploration, positioning itself as "Explorer of Future Mobility". Headquartered in Guangzhou, China, the company operates R&D centers in Beijing, Shanghai, Shenzhen, Zhaoqing, and Yangzhou, and has established intelligent manufacturing bases in Zhaoqing and Guangzhou.

XPENG pursues a global strategy for research, development, and sales, with an R&D center in the United States and subsidiaries across multiple European countries. The company adheres to full-stack in-house development of intelligent driver-assistance software and the development of core hardware, delivering an exceptional intelligent driving and riding experience for users.

On August 27, 2020, XPENG officially listed on the New York Stock Exchange (NYSE: XPEV), raising funds in an IPO that set a record at the time for the global new energy vehicle industry. On July 7, 2021, the company listed on the Hong Kong Stock Exchange (HKEX: 9868), becoming the first Chinese new-energy automaker to achieve dual primary listings in both Hong Kong and New York.

For more information, please visit https://www.xpeng.com/.

Contacts:
For Media Enquiries: Alison Liang, XPENG PR Department
Email: liangrq3@xiaopeng.com 

View original content:https://www.prnewswire.com/news-releases/xpeng-peking-university-collaborative-research-accepted-by-aaai-2026-introducing-a-novel-visual-token-pruning-framework-for-autonomous-driving-302650038.html

SOURCE XPENG

FAQ

AI-generated questions and answers. How Rhea-AI works. Not financial advice.

What did XPENG (XPEV) announce on December 29, 2025 about FastDriveVLA?

XPENG announced a AAAI 2026-accepted paper describing FastDriveVLA, a token pruning method that reduced tokens from 3,249 to 812 and cut computational load by ~7.5x on nuScenes.

How could FastDriveVLA affect XPEV vehicle autonomy performance and costs?

The framework targets lower onboard compute needs via 7.5x computational reduction, which may improve inference speed and hardware efficiency for autonomous stacks.

When was the XPENG–Peking University paper accepted to AAAI and how selective was it?

The paper was accepted to AAAI 2026; AAAI received 23,680 submissions with a 17.6% acceptance rate.

What benchmark and metrics validated FastDriveVLA for XPEV research?

FastDriveVLA was evaluated on the nuScenes benchmark, achieving state-of-the-art pruning performance while maintaining planning accuracy.

Does XPENG plan to use FastDriveVLA in production vehicles soon?

XPENG said the work reflects full-stack in-house capabilities and supports its L4 autonomy goals, but no commercial deployment timing was provided.