WORLDTECHPreprint links sparse autoencoder latents to route intent in self-driving planners
An arXiv preprint contends that end-to-end self-driving planners deliver strong results yet stay largely opaque, and offers a method for probing them from within. The authors state that unsupervised analysis of sparse autoencoder latents brings to light structure tied to driving behaviour.
What the authors report
The arXiv abstract describes end-to-end autonomous driving models as strong in planning but largely closed to inspection, a limitation the authors say weakens both observability and trust. Their unsupervised clustering of sparse autoencoder latents, they state, yields structure that matches route intent, vehicle speed and braking behaviour, and recovers route intent with 98.3% accuracy.
The preprint sits in the robotics category on arXiv and has been accepted to the Conference on Robot Learning (CoRL) 2026.
Steering latents shifts planner outputs
According to the authors, nudging individual latents produces changes in planner outputs that are predictable and monotonic, affecting lateral displacement, curvature and braking. Code and visualisations accompany the work.
Which claims stand on one source
Every result is the authors' own claim as set out in the arXiv abstract, and this report rests on that one source without independent checking. The abstract does not name the driving models, datasets or conditions behind the figures, so the accuracy and steering results can only be taken at the level the abstract states.
Open questions
- Which driving models and datasets produced the 98.3% accuracy figure?
- Do the steering effects hold across different driving scenarios?