AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

DriveMind: Mind-Evolving Belief Tracking for Closed-Loop Autonomous Driving

Yan Shi, Kehua Chen, Jiarui Qi, Yihong Tang, Yinhai Wang, Liao Haicheng

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DriveMind, a mind-evolving belief-tracking approach for MLLM-based closed-loop autonomous driving that maintains and revises hypotheses about the scene across frames instead of reasoning from scratch each frame.

Problem

MLLM-based autonomous drivers reason from scratch at each frame and commit to a single scene explanation, discarding earlier hypotheses, so under ambiguity they lock onto a wrong interpretation and act before evidence accumulates.

Approach

Maintains an evolving set of belief hypotheses about the traffic scene that is updated as evidence accumulates over frames, rather than committing to a single per-frame explanation. Full method details are not available in the provided text, so this is inferred from the abstract only.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add multi-hypothesis belief tracking (K=4) to the same VLM and diffusion controller (ablation)DriveMind · Bench2DriveDriving Score: 68.50 -> 81.37Improved
add multi-hypothesis belief tracking (K=4) (ablation)DriveMind · Bench2DriveSuccess Rate (%): 41.36 -> 54.09Improved
use 4 hypotheses instead of 2 (also adds deceleration and lane-change initial hypotheses) (ablation)DriveMind belief tracking · Bench2DriveDriving Score: 70.00 -> 81.37Improved
weight hypotheses by VLM consistency score instead of uniform weights (ablation)DriveMind belief weighting · Bench2Drive maneuver predictionManeuver accuracy: 0.779 -> 0.930Improved
weight hypotheses by VLM score instead of random weights (ablation)DriveMind belief weighting · Bench2Drive maneuver predictionManeuver accuracy: 0.644 -> 0.930Improved
increase hypothesis count K beyond 4 (ablation)DriveMind belief tracking · behavioral diversity (four scenarios)action disagreement / unique predictionsNo clear effect
use asynchronous slow-fast belief refresh instead of synchronous reasoning (ablation)DriveMind · Bench2Drive runtime (RTX A6000)control-step latency (ms): 177.0 -> 65.5Improved
increase hypothesis count K from 2 to 6 (ablation)High-level belief update · Bench2Drive runtimelatency (ms): 74.0 -> 165.5Hurt
add belief tracking (ablation)DriveMind · NAVSIMPDMS: 78.2 -> 81.0Improved
add belief trackingDriveMind · Bench2Drive Signalized IntersectionSuccess Rate (%): 12.0 -> 4.0Hurt
add belief trackingDriveMind · Bench2Drive Opposite DirectionSuccess Rate (%): 13.3 -> 13.3No clear effect
add belief trackingDriveMind · Bench2Drive Unsignalized IntersectionDriving Score: 54.4 -> 74.4Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy