EventOD: LLM-Guided OD Flow Generation for Disruptive Events
A team of researchers has introduced EventOD, a framework designed for generating event-adaptive origin-destination (OD) flows. This system utilizes a large language model (LLM) to derive functional and demographic control vectors at the regional level from broad event data. By employing structured event semantics, the framework directs a pretrained OD generator and incorporates two lightweight adaptation modules, AlphaNet and BetaNet, to adjust for semantic changes. Additionally, it features a retrieval-augmented fallback mechanism for instances of sparse supervision. This innovative method overcomes the limitations faced by deep OD models that are typically trained on standard mobility patterns when significant events disrupt regional functions and population dynamics, all without the need for retraining for each occurrence. The findings are available on arXiv under ID 2607.22655.
Key facts
- EventOD is an event-adaptive OD generation framework.
- It uses an LLM to infer region-level functional and demographic control vectors.
- It includes two adaptation modules: AlphaNet and BetaNet.
- It has a retrieval-augmented fallback pathway for sparse supervision.
- It steers a pretrained OD generator using structured event semantics.
- Existing deep OD models degrade under extreme events.
- Retraining a new generator for each event is impractical.
- Published on arXiv with ID 2607.22655.
Entities
Institutions
- arXiv