← News

ResearchSources: arXiv · 19 Nov 2026

DRIFT pairs a small language model with a switchable planner for autonomous driving

The arXiv paper describes a compact end-to-end driving framework whose language reasoning can be bypassed at runtime, cutting compute without changing driving results.

In plain words

Vision-language models can describe driving scenes, but they are too computationally heavy to run constantly. DRIFT uses a very small language model for scene descriptions and trains a planner that can also run on distilled summaries of that model's output. The planner can therefore skip the language model entirely when compute is limited.

What was announced

The authors say DRIFT integrates a lightweight language model, under 150 million parameters, that handles visual question answering and action commentary by projecting intermediate driving representations into the language embedding space. A planning module is trained on surrogate reasoning features distilled from this model and can be conditioned on either the language model or the surrogate features at inference, so deployments can bypass the language model without sacrificing driving performance, according to the paper.

On the Bench2Drive benchmark, the 255-million-parameter system reportedly reached a Driving Score of 97.03 and a Success Rate of 89.39%, averaged over three runs, by interleaving the language model and surrogate reasoning at a chosen rate. The authors describe this as on par with the strongest non-language models while being far more compact than existing vision-language approaches, and report competitive results on the real-world NAVSIM and AlpaSim benchmarks at significantly lower computational cost.

What it does not show

  • All results come from the paper's own abstract; no independent evaluation or confirmation is available, and the excerpt does not cover the full paper's methods or failure cases.

For security and IT teams

Scope
The source is the paper's abstract only, covering the framework design and headline benchmark numbers rather than implementation or deployment details.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.