DriveMind maintains weighted hypotheses. A vision-language model updates them; diffusion control generates waypoints from their blend.
The authors report Bench2Drive Driving Score rising from 68.50 to 81.37 and Success Rate from 41.36% to 54.09%, against their same-pipeline single-interpretation baseline.
The big change
What changed: An October 23, 2026 preprint (v1) reports gains from carrying weighted hypotheses through simulated driving.
Why it matters: Same-pipeline results probe belief tracking; broader initial maneuver coverage limits the four-hypothesis count ablation.
What to watch: The paper lists code as available; the repository says “Code coming soon.” BIG CHANGE did not run DriveMind; road safety and independent reproduction remain unestablished.
Sources & further reading
- DriveMind paper — Method and results.
- Project repository — Code status.
