AutoMine Solution for AV2 2026 Scenario Mining Challenge

AutoMine, a novel scenario mining method combining large language models and vision language models, achieved competitive scores in the Argoverse 2 Scenario Mining Competition at CVPR 2026. The approach addresses the critical challenge of extracting safety-critical scenarios from autonomous driving logs through self-refining code generation and execution feedback.

AutoMine tackles a fundamental challenge in autonomous vehicle development: extracting meaningful, safety-critical scenarios from massive datasets of driving logs for robust evaluation. The method combines semantic-preserving prompt augmentation to reduce LLM sensitivity with vision language models capable of handling perception noise, demonstrating how multi-modal AI systems can address complex real-world problems in autonomous systems.

The development of scenario mining techniques reflects broader industry trends in autonomous driving. As AV systems mature, data-driven evaluation becomes increasingly critical—companies cannot rely solely on synthetic testing. Real-world scenarios contain edge cases and complex interactions that simulators struggle to replicate. This competition validates the growing recognition that AI-powered data curation directly impacts AV safety and performance standards.

For the autonomous vehicle industry, robust scenario mining infrastructure offers significant competitive advantages. Engineers can systematically identify gaps in system performance, prioritize safety improvements, and establish standardized evaluation benchmarks. Winning solutions like AutoMine potentially shape how companies structure their testing pipelines and evaluate model readiness for deployment.

The technical approach—combining LLMs for semantic understanding, VLMs for visual reasoning, and execution feedback for refinement—establishes a repeatable framework other teams can build upon. As regulatory bodies increasingly demand transparent, data-driven safety validation, methods that efficiently extract meaningful scenarios from logs become essential infrastructure. The competition results suggest that AI-assisted scenario mining is transitioning from research novelty to practical necessity in AV development.

• →AutoMine combines LLMs and VLMs with execution feedback to efficiently mine safety-critical scenarios from autonomous driving logs.
• →The method achieved competitive scores (36.38 HOTA-Temporal, 77.21 Timestamp BA) in the CVPR 2026 Argoverse 2 competition.
• →Semantic-preserving prompt augmentation reduces LLM sensitivity, improving robustness across diverse driving scenarios.
• →Real-world scenario extraction is becoming essential infrastructure for AV safety validation and regulatory compliance.
• →Multi-modal AI approaches combining language and vision models prove effective for handling perception noise and open-world visual complexity.