Radar4D-VLM achieves 98% radar-only recall on frozen LLMs

Autonomous driving's vision-language models almost always lean on cameras and LiDAR, leaving 4D radar—which is robust to fog, rain, and darkness and directly measures radial velocity—as an afterthought. Radar4D-VLM changes that by treating radar as a standalone perception modality. Built by a team of academic and industry researchers, the model ingests ten consecutive 4D-radar point-cloud sweeps and organizes the data into a compact hierarchy of object, scene, and kinematic tokens. These tokens are fed through a parameter-efficient projector into frozen language backbones, meaning the language models are not fine-tuned at all. Auditable prediction heads then jointly estimate object count, spatial distribution, motion state, collision risk, semantic category, and radial velocity.

On the sequence-isolated K-Radar validation set, Radar4D-VLM's top-64 proposal recall hits 98.13% at a 4-meter threshold—outperforming fixed-lattice and uniform-random controls by 6.40 and 22.83 percentage points. The researchers tested 24 matched runs across five frozen language model families: Qwen, Phi, Mistral, Llama, and Gemma. The radar-token interface worked flawlessly with all of them, proving a universal compatibility layer. Interestingly, matched controls with aligned, permuted, and no-language supervision showed that language supervision didn't provide a stable, direct benefit to the prediction heads. That key result separates the value of a frozen-LLM interface from the value of language supervision, laying a reproducible foundation for radar-only scene and motion reasoning.