Research map / Vision-language and multimodal / Multimodal LLMs
Ask Less, See More: Communication-Conditioned Token Pruning for Vehicle-to-Vehicle Cooperative Autonomous Driving with Multimodal Large Language Models
Shiqi Sun, Yantao Lu, Bingkun Sun, Ning Liu, Bo Jiang, Ying Zhang, Jinchao Chen, Chenglie Du
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 communication-conditioned token pruning to reduce the communication and inference cost of V2V cooperative driving with MLLMs, based only on the abstract's stated problem framing.
Problem
Existing vehicle-to-vehicle (V2V) MLLM frameworks share and fuse dense tokens, causing high communication and inference costs for language-based cooperative driving in occluded scenarios.
Approach
Conditions token pruning on communication so that only informative tokens are shared and fused, moving beyond conventional feature-sharing V2V perception; details beyond the abstract are unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatial-redundancy term (lambda 1.2) to token scoring instead of lambda 0.0 (ablation) | V2V-CCM · V2V-QA 70% pruning | Q5 L2 error (m): 6.08 -> 5.21 | Improved |
| add spatial-redundancy term (lambda 1.2) instead of lambda 0.0 (ablation) | V2V-CCM · V2V-QA 70% pruning | Q5 collision rate (%): 7.54 -> 6.28 | Improved |
| increase lambda from 1.2 to 1.5 (ablation) | V2V-CCM · V2V-QA 70% pruning | Q5 L2 error (m): 5.21 -> 5.34 | Hurt |
| increase lambda from 0.5 to 1.0 (ablation) | V2V-CCM · V2V-QA 70% pruning | Q5 L2 error (m): 5.61 -> 5.43 | Improved |
| use communication-conditioned pruning instead of VisionZip at 70% | V2V-LLM · V2V-QA V2V-split 70% pruning | Q5 L2 error (m): 6.97 -> 5.21 | Improved |
| use communication-conditioned pruning instead of VisionZip at 70% | V2V-GoT · V2V-GoT-QA 70% pruning | Q9 L2 error (m): 5.48 -> 3.01 | Improved |
| use communication-conditioned pruning instead of Random at 30% | V2V-LLM · V2V-QA V2V-split 30% pruning | Q5 L2 error (m): 6.78 -> 5.05 | Improved |
| prune 30% of tokens with V2V-CCM | V2V-GoT · V2V-GoT-QA | Q9 L2 error (m): 2.62 -> 2.65 | No clear effect |
| prune 30% of tokens with V2V-CCM | V2V-LLM · V2V-QA V2V-split | grounding F1 (QGr): 40.7 -> 40.5 | No clear effect |
| prune 70% of tokens with V2V-CCM | V2V-LLM · V2V-QA V2V-split | Q5 L2 error (m): 4.99 -> 5.21 | Hurt |
| prune 70% of tokens with V2V-CCM | V2V-LLM · V2V-LLM efficiency | communication bandwidth (MB): 0.4068 -> 0.1251 | Improved |
| prune 70% of tokens with V2V-CCM | V2V-LLM · H200 GPU | TTFT (ms): 45.5 -> 22.0 | Improved |
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.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 7 of 7 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.
Closest papers by meaning
- Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving (CVPR 2026 · similarity 0.88)
- DriveGPT4-V2: Harnessing Large Language Model Capabilities for Enhanced Closed-Loop Autonomous Driving (CVPR 2025 · similarity 0.87)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Embodied Scene Understanding for Vision Language Models via MetaVQA (CVPR 2025 · similarity 0.87)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · similarity 0.87)
- Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models (ICML 2024 · similarity 0.86)
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.
