Personalizing public transit routing has long been a challenge because traditional routing algorithms struggle to incorporate the diverse preferences of individual travelers. A new framework called ChatPlanner, detailed in a paper posted on arXiv, addresses this by leveraging Large Language Models (LLMs) to interpret natural language queries and translate user preferences into routing objectives.
The research, conducted by Yang, Tingting, Xue, Chenhao, and Jun, proposes a framework that combines fine-tuned LLMs with Retrieval-Augmented Generation (RAG). According to the paper, ChatPlanner extracts routing parameters and interprets nuanced user preferences from natural language inputs, then integrates these preferences into the objective function of a public transit routing algorithm. The system was trained and evaluated on preference-aware datasets that incorporate eight personas and five contexts to establish scoring standards for both fine-tuning and RAG.
How ChatPlanner Works
ChatPlanner’s architecture consists of two core components: a fine-tuned LLM that learns general preference patterns and enforces the required output structure, and a RAG module that provides query-specific context. According to the paper, RAG helps resolve imprecise or conversational expressions and calibrates continuous scores, while fine-tuning ensures the model adheres to the expected format. Together, they achieve the highest accuracy in both routing information extraction and user preference interpretation.
The framework is designed to tackle the difficulty of capturing and integrating diverse user preferences into routing algorithms, a persistent challenge in public transit systems.
Experimental Validation
Three experiments were conducted to validate ChatPlanner’s capabilities. The first tested the feasibility of the solutions generated. The second focused on the extraction of routing information and preferences. The third evaluated the quality and completeness of the solution set. Results demonstrated that ChatPlanner reliably generates feasible solutions, according to the paper.
The table below summarizes the experiments and their outcomes:
| Experiment | Purpose | Outcome |
|---|---|---|
| Feasibility | Validate that generated solutions are practical | ChatPlanner reliably produces feasible routes |
| Information extraction | Assess accuracy in extracting routing parameters and preferences | Combined fine-tuning + RAG achieves highest accuracy |
| Solution set quality | Evaluate diversity and value of alternatives | Captures user preferences overlooked by existing planners |
Implications for Transit Optimization
By capturing user preferences, ChatPlanner identifies valuable solutions across different dimensions that existing route planners overlook, generating more valuable route alternatives. For example, a traveler who prioritizes minimal walking distance might receive a different route than one who values fewer transfers—something conventional planners often fail to offer.
The paper states that the integration of natural language understanding into transportation optimization establishes a new paradigm. For technology leaders in logistics and supply chain, the underlying principles—fine-tuned LLMs combined with RAG for preference extraction and integration into optimization algorithms—are directly transferable to problems like fleet routing, last-mile delivery, and intermodal transport planning. The framework demonstrates how LLMs can bridge the gap between human preferences and algorithmic optimization, potentially reducing planning time and increasing user satisfaction.
While the current study focuses on public transit, the approach could be adapted to any routing domain where user preferences are diverse and dynamic. The authors note that future work could expand to multi-modal transport or real-time dynamic preferences. As enterprises seek to personalize digital experiences, ChatPlanner offers a blueprint for combining generative AI with operational research.