CausalDriveBench: A Benchmark for Causal Reasoning in Autonomous Driving

815 samples across 475 scenes from the nuScenes dataset. The benchmark, created by causaldrivebench, bundles curated causal scene graphs with three types of multiple-choice and open-ended QA tasks. It was last updated on May 7, 2026.

Use Cases

• Evaluating causal reasoning models based on the provided active QA tasks.
• Training models for scene graph generation using the curated causal scene graphs.
• Benchmarking AI robustness against distractors using the distractor QA tasks.
• Studying latent causal relationships in driving scenarios with the dormant QA tasks.

Strengths

• 815 samples provide a substantial testbed for evaluation.
• Three distinct QA task types (active, dormant, distractor) allow for multi-faceted benchmarking.
• Built on the established nuScenes dataset, leveraging its real-world driving data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Built on top of the nuScenes autonomous driving dataset.

Collection Method

Likely involves curation and annotation of nuScenes data to create causal scene graphs and QA pairs.

Freshness

Last updated 2026-05-07 07:38:37; freshness should be verified.