Benchmarking Retrieval-Augmented Generation Strategies for Large Language Model-Based Travel Mode Choice Prediction
Researchers have developed a framework for using Large Language Models (LLMs) to predict travel mode choice. They tested four different retrieval strategies and found that one combination - using the GPT-4o model with balanced retrieval and cross-encoder re-ranking - achieved an accuracy of 80.8%. This is better than traditional statistical and machine learning models, and also shows superior zero-shot transfer abilities.
Researchers have developed a framework for using Large Language Models (LLMs) to predict travel mode choice. They tested four different retrieval strategies and found that one combination - using the GPT-4o model with balanced retrieval and cross-encoder re-ranking - achieved an accuracy of 80.8%. This is better than traditional statistical and machine learning models, and also shows superior zero-shot transfer abilities.
---
Why it matters: This matters because it demonstrates the potential for LLMs to improve travel behavior modeling, which is essential for effective transportation planning. The results show that aligning retrieval strategies with model capabilities can lead to significant improvements in predictive accuracy.
Source: https://arxiv.org/abs/2508.17527
This article was originally published at: https://arxiv.org/abs/2508.17527