Research map / Vision-language and multimodal / Multimodal LLMs
BEVLM: Distilling Semantic Knowledge from LLMs into Bird's-Eye View Representations
Thomas Monninger, Shaoyuan Xie, Qi Chen, Sihao Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BEV tokens instead of per-view image tokens as LLM input | InternVL3-1B · Ego3D object-centric | MCQ accuracy: 42.02 -> 61.34 | Improved |
| use BEV tokens instead of single-view image tokens | InternVL3-1B · DriveLM object existence | accuracy: 89.8 -> 90.8 | Improved |
| scale LLM from 1B to 8B | BEV tokens · DriveLM object existence | accuracy: 90.8 -> 95.3 | Improved |
| train MLP projector on BEV tokens (ablation) | InternVL3-1B on ViT tokens · DriveLM object existence | accuracy: 74.2 -> 90.3 | Improved |
| semantic distillation from LLM into BEV encoder (1B teacher) | UniAD · NeuroNCAP | score: 2.38 -> 2.93 | Improved |
| use 8B teacher instead of 1B (ablation) | BEVLM distillation on UniAD · NeuroNCAP | score: 2.93 -> 3.05 | Improved |
| transfer 8B distilled BEV encoder | VAD · NeuroNCAP | score: 3.24 -> 3.42 | Improved |
| distill from 1B teacher | UniAD · NeuroNCAP | collision rate: 0.56 -> 0.54 | No clear effect |
| distill from 8B teacher | UniAD · NeuroNCAP | collision rate: 0.56 -> 0.47 | Improved |
| distillation | UniAD · nuScenes open-loop | avg L2: 1.05 -> 0.97 | Improved |
| replace frozen LLM with random aux head on VQA (ablation) | UniAD · NeuroNCAP | score: 2.38 -> 1.75 | Hurt |
| VLM-AD-style CLIP text distillation instead of LLM (ablation) | UniAD · NeuroNCAP | score: 2.38 -> 2.07 | Hurt |
Closest papers by meaning
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.89)
- EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence (CVPR 2026 · similarity 0.89)
- Holistic Autonomous Driving Understanding by Bird's-Eye-View Injected Multi-Modal Large Models (CVPR 2024 · similarity 0.89)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.89)
- Multi-view Relational Distillation for Spatial Reasoning with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.89)
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.
