Research map / Vision-language and multimodal / Multimodal LLMs

DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Introduces DriveCombo, a CARLA-based multiple-choice benchmark (~70K questions) for compositional traffic-rule reasoning by MLLMs across a five-level cognitive ladder, plus a Rule2Scene agent that generates scenes from rule pairs; evaluates 14 MLLMs and shows LoRA SFT on the data helps.

Problem

Existing traffic-rule benchmarks for driving MLLMs focus on single rules and 2D static images, so they do not test whether models can reason over multiple concurrent or conflicting traffic rules in dynamic scenes.

Approach

A five-level cognitive ladder (single rule, multiple static rules, dynamic interactions, mixed static and dynamic, priority conflicts) built from atomic rules decomposed from country driving manuals, with rule pairs labeled by perceptual type and norm relation, and an LLM-plus-CARLA pipeline (Rule2Scene) that renders the composed rules into multi-frame driving scenes for visual MCQs.

Key results

• Proprietary and open models decline from L1 (up to 86.91% for GPT-5 pro) to L5 conflict resolution, where the best models reach only 41–44% accuracy against over 98% for human drivers.
• SFT with LoRA on DriveCombo raises Llama-3.2-11B's average accuracy by about 29.7 points and Qwen3-VL-8B's L5 accuracy to 60.2%.
• Fine-tuning InternVL-2.5-8B on DriveCombo lowers nuScenes average L2 error from 2.71 (DriveQA SFT) to 2.53 m.

Surprising findings

• Text-only descriptions (DriveCombo-Text) yield modestly higher accuracy than visual frames across all models, which the authors attribute to semantic loss in visual understanding.

Limitations

• Scene generation is constrained by CARLA's asset library.
• Evaluation is limited to multiple-choice questions with a zero-shot protocol and four frames per question.
• Human reference accuracy is based on only 30 drivers on 100 questions per level.

Experimental setup

Datasets: DriveCombo nuScenes DriveQA

Benchmarks: DriveCombo DriveCombo-Text nuScenes Bench2Drive

Metrics: accuracy L2 error driving score success rate

Baselines: DriveQA IDKB UniAD-Base VAD

Builds on: TARGET DriveQA Bench2Drive LLaMA-Factory LoRA

Closest papers by meaning

• DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.90)
• SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.90)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.90)
• MindDriver: Introducing Progressive Multimodal Reasoning for Autonomous Driving (CVPR 2026 · similarity 0.90)
• STSBench: A Spatio-temporal Scenario Benchmark for Multi-modal Large Language Models in Autonomous Driving (NeurIPS 2025 · similarity 0.90)
• nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving (NeurIPS 2026 · similarity 0.90)