To answer this question: Can vision-language models predict imminent vehicle collisions accurately, explicably and efficiently? Researchers at Imperial College London have developed a framework that combines task-specific fine-tuning, two-step chain-of-thought prompting and feature-based knowledge distillation. The customised model achieved 84.7% average accuracy while generating textual rationales for its predictions. After distillation, the compact model retained more than 95% of the teacher model’s prediction performance and reduced inference latency by over 50%.
The team published their study in Communications in Transportation Research ( https://doi.org/10.26599/COMMTR.2026.9640047 ).
Vehicle collision warning systems can support timely warnings and interventions, but conventional deep learning models often provide limited explanation for their predictions. Vision-language models offer a potential solution by jointly analysing driving videos and generating natural-language rationales. However, general-purpose models lack specialised collision-related knowledge and often require substantial computational resources.
To address these challenges, a general-purpose vision-language model was fine-tuned using real-world pre-collision and normal driving video clips. A two-step chain-of-thought prompt guided the model to first analyse the driving scene and identify potential risk sources, and then predict whether a collision was likely to occur imminently. The resulting model could therefore provide both a prediction and an accompanying explanation.
Knowledge was then transferred from the customised model to a smaller student model by aligning their intermediate hidden representations. Experimental results showed that the customised model achieved average accuracy, precision and recall of 84.7%, 83.6% and 86.8%, respectively, outperforming conventional deep learning baselines. The distilled model retained more than 95% of the teacher model’s prediction performance while reducing inference latency by over 50%.
The study provides a new approach to balancing prediction performance, explainability and computational efficiency in vision-based collision warning systems. Future research could incorporate additional vehicle kinematic information and further investigate the internal reasoning mechanisms of vision-language models.
DOI Link:
https://doi.org/10.26599/COMMTR.2026.9640047
About Communications in Transportation Research
Communications in Transportation Research was launched in 2021, with academic support provided by Tsinghua University and China Intelligent Transportation Systems Association. The Editors-in-Chief are Professor Xiaobo Qu, a member of the Academia Europaea from Tsinghua University, and Professor Xiaopeng (Shaw) Li from University of Wisconsin–Madison. The journal mainly publishes high-quality, original research and review articles that are of significant importance to emerging transportation systems, aiming to serve as an international platform for showcasing and exchanging innovative achievements in transportation and related fields, fostering academic exchange and development between China and the global community.
It has been indexed in SCIE, SSCI, Ei Compendex, Scopus, CSTPCD, CSCD, OAJ, DOAJ, TRID and other databases. It was selected as Q1 Top Journal in the Engineering and Technology category of the Chinese Academy of Sciences (CAS) Journal Ranking List. In 2022, it was selected as a High-Starting-Point new journal project of the “China Science and Technology Journal Excellence Action Plan”. In 2024, it was selected as the Support the Development Project of “High-Level International Scientific and Technological Journals”. The same year, it was also chosen as an English Journal Tier Project of the “China Science and Technology Journal Excellence Action Plan PhaseⅡ”. In 2024, it received the first impact factor (2023 IF) of 12.5, ranking Top1 (1/58, Q1) among all journals in "TRANSPORTATION" category. In 2026, its 2025 IF was announced as 12.7, maintaining the Top1 position (1/66, Q1) in the same category.
From Volume 6 (2026), Communications in Transportation Research will be published by Tsinghua University Press on the SciOpen platform with the official journal website at https://www.sciopen.com/journal/2097-5023 . We kindly request that all new manuscript submissions be made through the journal’s submission system at https://mc03.manuscriptcentral.com/commtr . For any submission-related inquiries, please contact the Editorial Office at commtr_e@mail.tsinghua.edu.cn.
Communications in Transportation Research
Distilling Vision-Language Models for Explainable Vehicle Collision Prediction
30-Sep-2026