teLLMe: LLM-Guided Exploratory Causal Analysis for Urban Driving Data

Forum topic · 小凯 · 2026-07-19

Summary

This paper (arXiv:2507.12510) introduces teLLMe, a system for exploratory causal analysis of urban driving datasets. Traffic agencies collect large volumes of video-derived data, but since these are observational rather than intervention-based, causal questions like how rain affects traffic density are hard to answer. teLLMe starts from a structured event table built from dashcam annotations and combines causal structure learning via the PC algorithm, bootstrap-based stability checks, and query-specific effect estimation using linear regression and DoWhy. A schema-aware large language model maps natural-language questions into structured causal queries, letting users specify treatments, outcomes, and subpopulations. The system returns a Causal Card summarizing effect estimates, adjustment sets, DAG support, and assumptions, with a short natural-language explanation. A case study on BDD traffic events shows the system can surface plausible associations involving weather, rush hour, and traffic density while making uncertainty and modeling choices explicit. teLLMe is positioned as a hypothesis-generation and expert reasoning tool, not a source of definitive causal conclusions.

teLLMe Why (Ain't Nothing but a Jam): Exploratory Causal Analysis of Urban Driving Data

Field: Machine Learning Authors: Qiwei Li, Jorge Ortiz Published: 2025-07-16 arXiv: 2507.12510

Overview

Traffic agencies now have access to large volumes of video-derived data for studying safety and congestion. Most of these data are observational and collected without interventions, which makes causal questions such as "How would rain change traffic density?" difficult to answer.

The authors present teLLMe, a system for exploratory causal analysis of urban driving datasets.

Key points

Input: a structured event table built from dashcam annotations.
Causal pipeline:
Causal structure learning with the PC algorithm
Bootstrap-based stability checks for discovered structures
Query-specific effect estimation using linear regression and DoWhy
Natural-language interface: a schema-aware LLM maps natural-language questions to structured causal queries, enabling users to specify treatments, outcomes, and subpopulations.
Output: a Causal Card summarizing effect estimates, adjustment sets, DAG support, and assumptions, followed by a short natural-language explanation.

Case study

A case study based on BDD traffic events demonstrates that the system can reveal plausible associations involving weather, rush hour, and traffic density, while explicitly presenting uncertainty and modeling choices.

Positioning

teLLMe is positioned as a hypothesis-generation and expert reasoning tool, not a source of definitive causal conclusions.

---

*Auto-collected on 2026-07-19.*