Go back to publications

WorldCorner: Evaluating Driving World Models for Safety-Critical Event Generation

Jiawei Wang*, Bodong Bi*, Kaiyang Luo, Xujie Song, Shaowei Hua, Haowei Sun, Jun Gao, Henry X. Liu† Submitted to AAAI 2027; under review
#World Models #Autonomous Driving #Benchmark #Safety
Project Page CodeComing soon CornerScenes DatasetComing soon

Abstract

Driving world models are increasingly evaluated on how real their output looks. For autonomous-driving safety, that is the wrong target: a generated clip can be photorealistic and still fail to reproduce the event it was asked for. WorldCorner is the first benchmark dedicated to evaluating the safety-critical event generation capabilities of driving world models, shifting the evaluation target from perceptual realism to event-level fidelity.

Contributions

CornerScenes

A curated dataset of over 1,200 safety-critical scenarios, built from adversarial modifications of real scenes. It covers two classes of interaction — those in which the ego vehicle acts as the adversarial actor, and those in which another vehicle does.

Benchmarking seven world models

Evaluating seven world models revealed a clear gap between conventional generation scores and safety-event fidelity:

  • Most models reached only 54%–66% event-progression accuracy.
  • Four of the seven scored below 40% on event-type accuracy.

In other words, models that score well on standard generation metrics frequently fail to produce the safety-critical event that was requested.

Context

Co-first author and project leader. I owned end-to-end experiment planning and execution, structuring the workflow into adversarial behavior generation and world-model rendering, and orchestrated 4,000+ GPU-hours of experiments across multiple compute nodes. The paper has been submitted to AAAI 2027 and is under review.

Advisor: Prof. Henry Liu, Mobility Transformation Lab, University of Michigan. Work carried out Mar 2026 – Jul 2026.