Understanding why traffic congestion happens requires more than just observing patterns in data. Researchers at Princeton University have developed a novel approach that combines large language models with causal inference techniques to help traffic agencies answer the kinds of "what if" questions that have historically been difficult to address.
The system, called teLLMe, represents a significant step forward in applying AI to real-world transportation challenges. Rather than simply correlating weather conditions with traffic density or identifying peak-hour patterns, the tool attempts to establish actual causal relationships: if it rained tomorrow, how would that meaningfully change traffic flow in a specific neighborhood?
According to arXiv, researchers Qiwei Li and Jorge Ortiz built teLLMe to process large volumes of dashcam footage and structured event annotations. The system converts unstructured questions posed in everyday language into formal causal queries that statistical models can analyze. A user might ask "Does rain cause more accidents on the highway during rush hour?" and the system translates that into a machine-readable format specifying the treatment variable (rain), the outcome (accidents), and the relevant subpopulation (highway drivers during peak times).
How It Works
The technical architecture layers multiple established techniques. The foundation is a structured event table extracted from dashcam annotations. The system then applies causal structure learning algorithms, specifically the PC algorithm, to identify potential causal relationships in the data. Bootstrap-based stability checks validate which relationships hold up under repeated sampling.
Where large language models enter the picture is in the interface layer. Instead of requiring users to write SQL queries or understand statistical notation, an LLM handles the translation between conversational English and structured causal queries. This schema-aware approach ensures the LLM respects the boundaries of what data is actually available and what relationships can meaningfully be tested.
Once the causal model is specified, teLLMe estimates effects using linear regression and integrates with DoWhy, a Python library for causal inference. The system then produces what researchers call a "Causal Card," a summary document that includes:
- Effect estimates with uncertainty intervals
- Sets of variables that need to be statistically controlled for
- Visual representations of the underlying causal assumptions
- Explicit discussion of modeling choices and their limitations
- Natural-language summary of findings
Designed for Exploration, Not Certainty
A critical design decision distinguishes teLLMe from systems that claim to deliver definitive answers. The researchers deliberately positioned it as a tool for hypothesis generation and expert reasoning rather than a source of ground truth. This reflects a mature understanding of how AI-assisted analysis should function in domains like traffic management, where policy decisions have real consequences.
Testing on traffic datasets derived from Berkeley DeepDrive data showed the system could surface plausible relationships involving weather conditions, time-of-day effects, and congestion patterns. The transparency about uncertainty and modeling choices appears particularly valuable for agencies that need to justify decisions to stakeholders.
As cities increasingly deploy connected infrastructure and cameras, the amount of observational traffic data available to transportation planners continues growing. TeLLMe demonstrates how combining LLMs with classical causal inference methods could help these agencies move beyond pattern recognition toward genuine understanding of traffic dynamics. The work also suggests a broader model for how large language models might enhance domain-specific analysis tools without replacing human expertise and judgment.
