KIDL: A knowledge-informed deep learning paradigm for generalizable and stability-optimized car-following models

As autonomous vehicles (AVs) integrate into real-world traffic, a key challenge remains: how to ensure their car-following behavior is both human-like and traffic-stable. Traditional car-following models (CFMs), whether physics-based or data-driven, often fail to generalize across diverse driving scenarios or maintain traffic flow stability.

 

This research, published on 20 September 2025, in Communications in Transportation Research, is a collaborative effort among Xi’an Jiaotong-Liverpool University, Monash University, University of Liverpool, and City University of Hong Kong, integrating expertise in transportation engineering, machine learning, and intelligent systems.

 

The study introduces KIDL—Knowledge-Informed Deep Learning—a groundbreaking framework that distills expert knowledge from Large Language Models (LLMs) into a lightweight and stability-optimized neural network for AV control. Unlike prior models limited by dataset biases, KIDL captures generalized driving principles by querying LLMs for thousands of diverse traffic scenarios and refining predictions through a structured knowledge distillation process.

 

KIDL’s innovation lies not only in its strong behavioral generalization but also in its theoretical guarantees: by incorporating local and string stability constraints into training, it ensures smooth and safe vehicle interactions in mixed traffic. Experiments on real-world datasets (NGSIM, HighD) confirm KIDL outperforms physics-based, deep learning, and hybrid models in both accuracy and safety—with zero collisions and superior generalization.

 

KIDL represents a new paradigm in autonomous vehicle design, blending human-like reasoning with formal stability control—scalable, interpretable, and ready for real-world deployment.

 

Car-following models (CFMs) are microscopic traffic models that capture longitudinal interactions between leading and following vehicles. Most CFMs follow a model-centric design and are calibrated or trained on specific datasets. While this yields high accuracy within seen scenarios, performance often degrades under unseen conditions due to the out-of-distribution generalization problem. Individual datasets rarely capture the full range of real-world variability, limiting model robustness. Although data-centric approaches that focus on collecting broader datasets can improve generalization, they are costly and difficult to scale. At the same time, ensuring traffic flow stability has become increasingly important for the deployment of CFMs in AV-integrated traffic systems, where safety and efficiency are critical priorities. This highlights the need for a new modeling paradigm that simultaneously addresses the limitations of generalization and the rising demand for stability in mixed traffic environments. To this end, we propose a Knowledge-Informed Deep Learning (KIDL) paradigm that jointly enhances behavioral generalization and traffic flow stability.

 

KIDL improves generalization by leveraging large language models (LLMs) as knowledge sources that encapsulate high-level car-following behaviors learned from diverse textual data, including driving rules, safety guidelines, and human reasoning. This enables KIDL to capture generalizable principles that extend beyond the scope of any single traffic dataset. Through knowledge distillation, insights from LLMs (as teachers) are transferred to lightweight neural networks (as students), forming a compact and efficient representation. Rather than employing LLMs as end-to-end models, KIDL adopts a distillation-based approach with three key advantages.

 

The first advantage is computational efficiency. LLMs generate linguistic responses sequentially and require substantial memory and processing resources, making them unsuitable for real-time applications. In contrast, KIDL produces single-step numerical predictions with significantly fewer parameters, enabling real-time inference at a fraction of the computational cost. The second advantage is the prediction reliability. LLMs may produce inaccurate or unfaithful content, which poses serious risks in safety-critical contexts. KIDL reduces this risk by applying self-consistency with majority voting during knowledge extraction, improving reliability and minimizing the likelihood of erroneous behavior.

 

The third advantage is theoretical tractability. The black-box nature, complex architectures, and dependence on natural language inputs and outputs make LLMs difficult to interpret and unsuitable for formal analysis, limiting their applicability in stability studies such as local and string stability. By distilling knowledge into a simplified surrogate model with numerical inputs and outputs, KIDL enables interpretable and analytically tractable stability analysis.

This property further allows KIDL to incorporate physically grounded stability constraints directly into the training objective, ensuring compliance with both local and string stability conditions. As a result, the model suppresses disturbance amplification and promotes smooth traffic flow.

 

By integrating behavioral fidelity with stability optimization, KIDL provides a scalable and robust solution for deployment in mixed traffic environments. This combination of generalizable behavior modeling and explicit stability assurance addresses a critical gap between human driver emulation and control-oriented AV deployment. To the best of our knowledge, KIDL is among the first frameworks to systematically achieve both objectives within a unified paradigm by distilling car-following knowledge from LLMs into a stability-aware neural architecture.

 

The above research is published in Communications in Transportation Research (COMMTR), which is a fully open access journal co-published by Tsinghua University Press and Elsevier. COMMTR publishes peer-reviewed high-quality research representing important advances of significance to emerging transport systems. COMMTR is also among the first transportation journals to make the Replication Package mandatory to facilitate researchers, practitioners, and the general public in understanding and advancing existing knowledge. At its discretion, Tsinghua University Press will pay the open access fee for all published papers in 2025.

 

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 Shuai’an Wang from Hong Kong Polytechnic University. 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 Phase Ⅱ”. 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 2025, its 2024 IF was announced as 14.5, maintaining the Top 1 position (1/61, Q1) in the same category. Tsinghua University Press will cover the open access fee for all published papers in 2025.

---

---