LLM-Driven Trajectory Prediction at Intersections with Chain-of-Thought Reasoning

Accurate vehicle trajectory prediction at signalized inter-sections is critical for autonomous driving safety, yet remains challenging due to complex vehicle-signal interactions and stringent real-time requirements. This paper proposes chain-of-thought (CoT)-intersection vehicle trajectory predict, a novel framework that integrates CoT prompting with lightweight knowledge distillation to achieve high-precision, explainable trajectory forecasting. We use structured CoT prompts to explicitly encode traffic rules and right-of-way logic into GPT-4 Turbo's reasoning process, mitigating rule agnosticism, and we employ dynamic knowledge distillation to transfer LLM capabilities into a compact Qwen-1.5 model. The framework employs a causal spatio-temporal fusion network that integrates distilled semantic features with kinematic data while preventing future information leakage. Evaluated on our newly released SIGNAL, CoT-intersection vehicle trajectory predict achieves: i). 0.93 m Average Displacement Error (ADE) and 1.86 m Final Displacement Error (FDE), outperforming state-of-the-art baselines like V2X-TrajNet by 23.1% and 13.5%, respectively; ii). 1.7% Signal Violation Rate (SVR) and 0.12 Conflict Severity Index (CSI), demonstrating 97.3% regulatory compliance and 72.1% collision risk reduction; iii). 28 ms inference latency on NVIDIA Jetson Orin, enabling real-time edge deployment. These advancements demonstrate the viability of lightweight, explainable LLM surrogates for safety-critical autonomous driving systems.

• Publication year

  2025

• Venue

  IEEE International Conference on Network Infrastructure and Digital Content

• Publication date

  2025-11-07

• Fields of study

  Not labeled

• Identifiers
  DOI 10.1109/IC-NIDC67200.2025.11390125
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• WAKE: Towards Robust and Physically Feasible Trajectory Prediction for Autonomous Vehicles With WAvelet and KinEmatics Synergy

  2025cited by this paper
• Physics-Informed Neural Network Model for Predictive Risk Assessment and Safety Analysis

  2024cited by this paper
• LLaMA: Open and Efficient Foundation Language Models

  2023cited by this paper
• A Comprehensive Survey of ChatGPT: Advancements, Applications, Prospects, and Challenges.

  2023cited by this paper
• Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving

  2023cited by this paper
• GPT-4 Enhanced Multimodal Grounding for Autonomous Driving: Leveraging Cross-Modal Attention with Large Language Models

  2023cited by this paper
• DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models

  2023cited by this paper
• Traffic Management: Multi-Scale Vehicle Detection in Varying Weather Conditions Using YOLOv4 and Spatial Pyramid Pooling Network

  2022cited by this paper
• Autonomous Traffic System for Emergency Vehicles

  2022cited by this paper
• Intention-Aware Vehicle Trajectory Prediction Based on Spatial-Temporal Dynamic Attention Network for Internet of Vehicles

  2022cited by this paper
• Short-term traffic flow prediction based on improved wavelet neural network

  2020cited by this paper
• Deep and Embedded Learning Approach for Traffic Flow Prediction in Urban Informatics

  2019cited by this paper
• Random Erasing Data Augmentation

  2017cited by this paper
• Attention is All you Need

  2017cited by this paper
• Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation

  2014cited by this paper
• Language Models

  2009cited by this paper

• No citing papers are available for this paper.