Driving the Wrong Way: Interpretability in End-to-End Driving

Franz Motzkus and Sebastian Bernhard submitted Driving the Wrong Way to arXiv on 7 Jul 2026 (arXiv:2607.06328), proposing a post hoc interpretability module that decomposes end-to-end driving behavior into semantically meaningful concepts and then uses those concepts to manipulate and correct trajectory predictions. The submission PDF is 24,149 KB and the paper links to an arXiv DOI.

What did Motzkus and Bernhard do?

They integrated unsupervised dictionary learning as a post hoc interpretability module into state-of-the-art end-to-end driving models and demonstrated that the extracted concepts causally influence driving decisions. The paper presents a stepwise framework that extracts and interprets meaningful concepts, connects them to multifaceted model outputs, and shows that targeted interventions at the concept level can alter and improve predicted future trajectories.

The authors frame the work as a way to reduce model opacity and uncover erroneous behavior inside complex driving models. The abstract states the approach both reveals decision-making logic for trajectory prediction and enables targeted mitigation that yields measurable improvements in overall driving performance.

How does the interpretability module work?

The core idea is post hoc unsupervised dictionary learning that decomposes the model's internal representations into semantically meaningful concepts, then links those concepts to the model's predicted future trajectories. The paper describes a stepwise framework: extract concepts with unsupervised dictionary learning, interpret those concepts against the model outputs, and perform targeted interventions at the concept level to test causal effect on driving decisions.

The module sits after a trained end-to-end model and produces a set of interpretable concept components. By intervening on those components the authors demonstrate manipulation of driving outputs, i.e., corrected trajectory predictions. The work emphasizes causal influence: concepts are not merely descriptive but are shown to change model behavior when altered.

What evidence and claims are in the paper?

The submission characterizes the contributions as threefold: (1) introducing unsupervised dictionary learning as a post hoc interpretability module for driving models, (2) providing a stepwise framework that connects extracted concepts to multifaceted model outputs for future-trajectory prediction, and (3) showing that targeted, concept-level interventions can manipulate and correct driving decisions, producing measurable improvements in overall driving performance. The abstract explicitly states these outcomes without giving numeric benchmarks in the submission metadata.

The authors position the method as a way to reduce opacity and to uncover erroneous or undesired behavior learned by end-to-end systems. They claim the approach enables targeted mitigation that in turn boosts model performance.

Why it matters

End-to-end driving models are increasingly complex and opaque, and this paper offers a concrete technique to translate internal representations into human-interpretable concepts and to test their causal role. If the concept extraction and intervention pipeline is robust, it gives practitioners a targeted lever to correct specific failure modes without retraining entire models. That could change how teams debug and harden end-to-end systems by shifting some fixes from black-box retraining to concept-level correction.

What to watch

Look for follow-up evaluations and shared code or data that reproduce the paper's claim that concept-level interventions lead to measurable improvements in driving performance. Confirmation will require independent tests on standard driving benchmarks or open datasets and clearer quantitative results beyond the submission abstract.

Bibliographic note: the paper is arXiv:2607.06328, submitted 7 Jul 2026, authors Franz Motzkus and Sebastian Bernhard, PDF size 24,149 KB.