teLLMe: LLM-Assisted Exploratory Causal Analysis for Urban Driving Data (arXiv 2507.12510)

Forum topic · 小凯 · 2026-07-19

Summary

teLLMe is a system for exploratory causal analysis of urban driving datasets, presented in arXiv paper 2507.12510 by Qiwei Li and Jorge Ortiz (July 2025). Traffic agencies hold large volumes of observational, video-derived data collected without interventions, making causal questions like "How would rain change traffic density?" hard to answer. teLLMe starts from a structured event table built from dashcam annotations and combines causal structure learning with the PC algorithm, bootstrap-based stability checks, and query-specific effect estimation using linear regression and DoWhy. A schema-aware LLM maps natural-language questions to 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, plus a brief natural-language explanation. A case study on BDD traffic events shows plausible associations involving weather, rush hour, and traffic density, with uncertainty and modeling choices made explicit. teLLMe is positioned as a hypothesis-generation and expert reasoning tool rather than a source of definitive causal conclusions.

Paper Overview

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

What the Paper Does

Traffic agencies now have access to large volumes of video-derived data for studying safety and congestion. However, 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.

teLLMe is a system for exploratory causal analysis of urban driving datasets.

How It Works

Builds a structured event table from dashcam annotations
Performs causal structure learning with the PC algorithm, plus bootstrap-based stability checks
Runs query-specific effect estimation using linear regression and DoWhy
Uses a schema-aware LLM to map natural-language questions to structured causal queries, so users can specify treatments, outcomes, and subpopulations

Output

teLLMe returns a Causal Card summarizing:

Effect estimates
Adjustment sets
DAG support
Assumptions
A short natural-language explanation

Case Study

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

Positioning

The authors position teLLMe as a hypothesis-generation and expert reasoning tool, not a source of definitive causal conclusions.

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