Data-Centric Post-Training for Improving Long-Tail Scenario Generalization in End-to-End Autonomous Driving

Abstract

End-to-end autonomous driving has advanced rapidly by integrating perception, prediction, and control into unified learned systems. Yet deployment remains constrained by long-tail scenarios: rare, ambiguous, or geographically specific situations that are underrepresented in training corpora but disproportionately important for safety. This paper argues that improving long-tail generalization requires a data-centric post-training paradigm rather than isolated architectural refinement. Post-training, in this context, refers to the structured adaptation of a pre-trained driving policy using curated, targeted, and often synthetic data, together with human feedback, simulation, and validation loops. The paper examines the structural trade-offs that arise when data-centric post-training becomes a system-level strategy. It considers how data infrastructure, labeling governance, scenario mining, simulation fidelity, fleet telemetry, and safety assurance interact across the autonomous driving stack. It also explores deployment consequences, including robustness under distribution shift, fairness across regions and road users, sustainability of continual learning pipelines, and policy implications for certification and liability. The central claim is that long-tail performance is not merely a model property but an institutional and infrastructural achievement. Effective post-training therefore depends on governance mechanisms that make rare events discoverable, comparable, auditable, and reusable without creating brittle overfitting to narrow test distributions. The paper concludes that data-centric post-training offers a credible path toward safer end-to-end autonomy, but only if technical pipelines are coupled with transparent data provenance, operational design domain discipline, and cross-organizational evaluation practices.