ChatPlanner: LLM framework for personalized transit routing

ChatPlanner introduces a framework that applies large language models to personalized public transit routing. Submitted to arXiv on 13 Jun 2026, the paper describes a pipeline that uses fine-tuned LLMs together with Retrieval-Augmented Generation to extract routing parameters and interpret nuanced user preferences, then incorporates those preferences into the objective function of a transit routing algorithm.

How does ChatPlanner work?

ChatPlanner uses fine-tuned LLMs combined with Retrieval-Augmented Generation to turn natural language queries into routing parameters and calibrated preference scores, and then injects those preferences into the optimization objective of a public transit routing algorithm. The system relies on fine-tuning to enforce output structure and learn general preference patterns, while RAG supplies query-specific context to resolve conversational or imprecise expressions and calibrate continuous scores.

The authors frame the pipeline as two complementary capabilities. Fine-tuning shapes the model output so routing information is extractable and consistent. RAG augments the LLM with external context at query time, which the paper says helps interpret ambiguous user comments and produce calibrated numeric scores that the routing algorithm can use.

What experiments did the authors run and what did they find?

The paper reports three experiments using preference aware datasets that incorporate eight personas and five contexts to evaluate feasibility, extraction accuracy, and the quality and completeness of generated solution sets. Results show ChatPlanner reliably generates feasible route solutions, with fine-tuning enforcing required output structure and learning general preference patterns, RAG supplying query-specific context and calibrating continuous scores, and the two combined achieving the highest accuracy in routing information extraction and user preference interpretation.

The study also includes selected case studies where capturing user preferences allowed ChatPlanner to identify route alternatives across different dimensions that existing planners overlook. The authors present these findings as evidence that integrating natural language understanding uncovers valuable solutions beyond standard route planners.

Why it matters

ChatPlanner links conversational preference expression to algorithmic routing decisions, offering a concrete method to make transit planners preference aware. The combination of fine-tuning and RAG addresses both generalization across users and specificity for individual queries, which matters because transit choices often hinge on heterogeneous and subtle preferences that traditional planners do not capture. If adopted, this approach could change how planners present alternatives, prioritizing routes that better match individual tradeoffs.

What to watch

The paper is listed as Under Review at Transportation Research Part C; acceptance there would be the next formal milestone. The arXiv entry also lists associated code, data, demos and hosting toggles such as Hugging Face Spaces and Replicate in its metadata, so look for published datasets, demonstration spaces or code releases tied to the submission that would enable reproduction and wider testing.

References and concrete facts from the submission: the manuscript titled "ChatPlanner: A Large Language Model Framework for Personalized Public Transit Routing" was submitted to arXiv on 13 Jun 2026 as arXiv:2606.15315, uses fine-tuned LLMs with Retrieval-Augmented Generation, defines preference aware datasets with eight personas and five contexts, and reports three experiments; the submission notes it is Under Review at Transportation Research Part C.