Researchers have introduced DriveDNA, a large-scale multimodal naturalistic dataset designed to benchmark personalized driving-style identification. The dataset comprises 975 hours of human-controlled driving, captured at 10 Hz, featuring 4,121 drives from 465 drivers across 115 different vehicle models. To facilitate complex analysis, the authors provided behavioral annotations alongside 276,248 rule-generated maneuver events spanning six classes, many of which underwent large-scale human auditing. The benchmark evaluates driving style through three distinct tasks: few-shot driver re-identification, personalized behavior prediction, and condition-matched comparison. Experimental results indicate that learned representations significantly outperform classical descriptors in identifying unseen drivers, achieving an AUROC of 0.935 compared to 0.707. Notably, while video-only models achieved comparable re-identification accuracy, they suffered from severe route leakage, suggesting they rely on contextual shortcuts rather than actual driving behavior. Conversely, learned representations maintained driver-specific information even when driving conditions were matched, whereas descriptor performance dropped to near-chance levels. These results emphasize the necessity of evaluating driving-style models based on their ability to isolate driver-specific behavioral patterns from confounding variables such as vehicle type, specific routes, and environmental conditions.