Claude Code-Driving Scenario Mining for the Argoverse 2 Challenge

A research team has submitted a four-stage pipeline to the CVPR 2026 Argoverse 2 Scenario Mining Challenge, using a large language model agent to autonomously generate code for mining driving scenarios. [1] The system, described in a paper posted to arXiv on June 8, 2026, relies on a Claude Code agent powered by GLM~5.1 to produce code without direct human authorship. [1] That code is then refined through a sequence of automated steps designed to improve reliability and reduce false detections. [1] In the second stage, the pipeline screens an iterative training set using a Timestamp Balanced Accuracy threshold of 0.8 to curate few-shot examples. [1] A separate Claude Code session then performs a semantic review of the generated code, checking for logical consistency and alignment with the challenge criteria. [1] The final stage applies Qwen3-VL, a vision-language model, to verify scenes and filter false positives before final submission. [1] The work targets the Argoverse 2 Scenario Mining Challenge, a competition held alongside the Computer Vision and Pattern Recognition conference. The challenge tasks participants with identifying safety-relevant driving scenarios from the Argoverse 2 dataset, a large-scale collection of real-world driving data. The authors report results on the official Argoverse 2 test set, though specific performance metrics beyond the pipeline description were not detailed in the abstract. [1] Automated scenario mining has become a focus area for autonomous vehicle development, where engineers seek to surface rare but critical events — such as near-misses or sudden braking — from thousands of hours of logged driving data. Manual review of such volumes is impractical, pushing teams toward machine-learning pipelines that can triage and rank candidate scenes. The use of an autonomous coding agent to build the mining logic itself represents a further step in automating the development workflow. [1]