Key hallucination types common in the transportation domain typically align with broader natural language generation hallucinations but have unique manifestations related to transit and traffic data. They include:
- Factual Hallucinations:
The model generates false or fabricated transportation facts, such as incorrect traffic incident reports, wrong vehicle counts, mishandled route schedules, or invented infrastructure details not supported by real data. - Contextual or Faithfulness Hallucinations:
Output contains information inconsistent with the transportation context or source data; for example, substituting the wrong road names, accident locations, or times despite partial matching cues. - Detached or Out-of-Distribution Hallucinations:
Details are irrelevant or unrelated to the transportation scenario prompted, such as adding unrelated geographic locations or non-transport entities that confuse critical situational awareness. - Attribute-Level Hallucinations:
Errors in characteristics specific to transit elements—like bus speed, vehicle types, or weather conditions linked to traffic incidents—where basic facts are correct but finer details are misspecified. - Temporal and Spatial Hallucinations:
Incorrect temporal or spatial relationships within the transportation data arise, such as reversing event sequences, misplacing incidents geographically, or mixing time zones and dates.
Understanding and categorizing these forms helps tailor annotation and mitigation strategies specific to transport analytics and smart city applications where accuracy and safety are paramount. Models often struggle most with low-data or ambiguous context situations, increasing hallucination propensity in rare or exceptional traffic scenarios.