Counterfactual Traffic Forecasts as Fact-Checkers for LLM Incident Reports

Operators increasingly ask language models to explain what happened on a road network: "the crash on the northbound freeway at 8:10 caused slowdowns three exits upstream that cleared by 9:30." Such sentences are easy to write and hard to check. The claim is about cause. It says the slowdown would not have happened without the crash, and a lookup in the sensor database cannot confirm that, because the database only records what did happen. We propose to check these claims with a spatio-temporal graph neural network, a model that forecasts sensor readings by passing information along the road graph and across time. The network is trained only on incident-free traffic and used as a counterfactual: it predicts what each sensor would have read had the incident not occurred. The gap between observed and predicted traffic, spread over the graph, gives an estimated "incident footprint." Each where, when, and how-long element of the LLM's explanation is then scored against that footprint.

The idea turns a known weakness into a tool. Standard traffic forecasters fail sharply during accidents and closures. TraffiDent and XTraffic call this the "Incident Gap." We treat that failure as a measurement: where a well-calibrated normal-conditions model is badly wrong, something abnormal is happening. The payoff is a verifier that catches spatio-temporal hallucinations that language models demonstrably cannot catch themselves. It gives localized feedback the LLM can act on, and it says "unverifiable" when the sensors cannot settle the question.

Core Claim

Our claim is that a graph-based counterfactual forecaster detects hallucinated spatial extent, onset time, and duration in LLM-written incident explanations more reliably than four alternatives. The first is text retrieval. The second is LLM self-judgment. The third is a tool-using LLM that compares against historical averages. The fourth is the same verifier with the road graph removed. We expect the advantage to be concentrated on claims about multi-hop propagation, meaning congestion that spreads to sensors several road segments away, and on days whose history is irregular.

The motivating evidence comes from benchmarks. Li et al.'s traffic-incident study evaluates nine foundation models on Vienna transit disruption reports. It finds 22% accuracy on "where" and 5.5% on "when," and finds that retrieval fixes invented station names but not temporal or relational reasoning. STBench shows that LLMs treat coordinates as symbols: open-source models score near or below 15% on direction tasks, and trajectory errors exceed 150 meters even after fine-tuning.

The closest methods are graph-based verifiers of LLM output. GNNVerifier converts tool-use plans into static dependency graphs and scores them with a graph network. It shows that LLM critics are fooled by fluent narration. GUARDIAN builds a temporal graph over agent conversation rounds and uses reconstruction error to prune hallucinating agents. Our proposal differs from both in what the graph represents. In both, the graph is built from the LLM's own output, whether plan steps or messages. In ours, the graph is the physical system the text describes, and the verifier's reference is a forecast of that system under a "no incident" hypothesis.

Other work fuses graph embeddings into the LLM. UrbanGPT feeds temporal-convolution features to an LLM for zero-shot forecasting. TGL-LLM feeds temporal-graph embeddings to an LLM for event multiple-choice questions. Neither checks free-text claims. UrbanGPT also deliberately discards the road graph, and our graph-removal ablation tests whether that choice hurts verification.

Research Program

The system has three parts.

The first is claim extraction. The LLM receives an incident record, the surrounding sensor time series, and the road graph. It writes an explanation under a constrained output schema. Each explanation is parsed into typed claim elements: affected segments or sensors, onset time, clearance time, direction of spread, and magnitude of slowdown. Parse failures are logged and counted as their own error category, not silently dropped.

The second is the counterfactual forecaster. We train a decoupled spatio-temporal attention model in the style of GMAN on windows with no incident within a buffer of the graph neighborhood. GMAN combines attention over nodes, attention over time, and direct history-to-future cross-attention, and it stays stable when 90% of its inputs are dropped. That robustness matters because incident-adjacent sensors are often noisy. The forecaster outputs quantiles or ensemble spread, which gives a calibrated predicted value $\hat{x}_{i,t}$ and scale $\sigma_{i,t}$ for sensor $i$ at time $t$, conditioned only on data from before the incident. The footprint is the set of sensor-times whose standardized deviation exceeds a threshold $\tau$:

$$z_{i,t} = \frac{x_{i,t} - \hat{x}_{i,t}}{\sigma_{i,t}}, \qquad F = \{(i,t) : |z_{i,t}| > \tau\},$$

where $x_{i,t}$ is the observed reading. Calibration is checked on held-out normal days, so that $\tau$ has a known false-alarm rate. We keep only the connected component of $F$ on the road graph that touches the incident location. This spatial connectivity requirement is the one place the graph acts as a hard constraint, and it separates incident impact from unrelated anomalies elsewhere.

The third is element-level scoring and repair. Each claimed element is compared with the footprint. Onset and clearance claims are compared with the first and last footprint times at the claimed sensors. Spatial claims are compared by overlap between the claimed and footprint sensor sets. Each element receives a risk score. An element touching sensors with missing data, or segments with no sensor, is labeled unverifiable instead of true or false. Following GNNVerifier's use of localized risk to request surgical edits, the high-risk elements and their footprint evidence are returned to the LLM for one round of targeted revision.

Baselines

The naive reference point is to accept the LLM's explanation as written. This gives the raw hallucination rate against which every detector is measured. Li et al. supplies the only reported comparable numbers: 22% spatial and 5.5% temporal accuracy on Vienna transit incidents. Those numbers come from a different network and task format, so we treat them as context, not as a target.

The strongest competitor is a tool-using LLM verifier. The same LLM, or a stronger one, gets a code interpreter and the full sensor history. It is asked to verify each claim element, typically by comparing against same-time-of-day, same-weekday averages. This is the approach a practitioner would build first, and it attacks our core premise directly. It gets the same data, the same claim schema, and a matched budget of tool calls and tokens. We will not hand-tune its prompts more than we tune our thresholds.

The key ablations share our pipeline and change only the counterfactual source. One replaces the forecaster with a seasonal historical average. One replaces it with a graph-free temporal convolution encoder of the UrbanGPT type. Both are trained on identical incident-free data with the same hyperparameter search budget. If the historical-average variant matches the full system, the forecaster adds nothing, and we will report that.

We also run two text baselines. One is an LLM judge given the raw data, the setting GNNVerifier found to be fooled by fluent narration. The other is retrieval of similar past incident records into the prompt.

Beating a baseline means a statistically significant gain in element-level detection AUROC on the semi-synthetic set, with confidence intervals from paired bootstrap. It also means a gain in footprint overlap after repair on real incidents. There is no reported result on Lacuna for hallucination detection on XTraffic or TraffiDent incident explanations, so every baseline is run by us.

Why not just…?

Why not just let a tool-using LLM query the sensor data? This is the real threat. For simple claims like "speeds at sensor 12 dropped at 8:20," a direct query will likely suffice, and we expect to tie there. The case for a forecaster rests on two kinds of claims. The first is causal: "because of the crash" requires a counterfactual that seasonal averages get wrong on irregular days. The second is propagation: "spread three exits upstream" requires distinguishing incident congestion from ordinary recurrent congestion several hops away. T-GCN reports that joint graph-temporal models beat historical average and purely temporal models. It cuts RMSE by about 3.3% over a GRU and 57.8% over a spatial-only GCN on SZ-taxi at 15 minutes, which suggests graph structure gives better normal-conditions references. But that is a forecasting result, not a verification result. Our experiments stratify by hop distance and by day regularity, so the settling test is direct: if the tool-using LLM matches us on multi-hop and holiday strata, the proposal fails.

Why not just retrieve past incidents into the prompt? Li et al. found that retrieval stops invented entities but does not fix temporal or duration reasoning. Duration is precisely the element we target. We include retrieval as a baseline to confirm this on freeway data.

Why not just fine-tune the LLM on incident explanations? STBench shows that fine-tuning lifts categorical tasks dramatically, from 19.89% to 91.98% on administrative region determination for Gemma-2B. It leaves metric tasks broken, with trajectory errors still above 150 meters. Onset and duration are metric quantities. Fine-tuning also yields a better writer, not a checker, and it would still need our verifier to measure its errors.

Why not just ask the LLM to judge itself? GNNVerifier reports that LLM critics miss structural errors hidden behind plausible narration, and GUARDIAN reports that agents amplify peers' errors. We include the judge baseline anyway, since traffic incidents are a new domain.

Proposed Experiments

We build on XTraffic and TraffiDent. Both align Caltrans PEMS sensors on a 16,972-node graph with timestamped incident records across seven classes.

The first experiment, and the one we treat as decisive, is semi-synthetic, because real footprints have no clean ground truth and we must avoid grading the verifier with its own output. We inject speed drops into real incident-free traffic, with known location, onset, duration, and upstream spread, using propagation patterns fitted to real incidents. We then collect LLM explanations and label each claim element against the injected truth. We report element-level AUROC and F1 for each verifier, stratified by hop distance from the incident and by day regularity. We predict the full system's margin over the historical-average ablation grows with hop distance and on holidays and weekends. A flat margin would falsify the graph's role.

The second experiment uses real incidents. The recorded incident start and end times serve as weak labels for onset and clearance claims. On a sample of several hundred incidents, two annotators label spatial extent independently from the raw plots, and we report their agreement. Metrics are detection AUROC, onset and clearance error in minutes, and spatial overlap between claimed and annotated sensor sets.

The third experiment tests repair. After one round of feedback, we measure the drop in hallucinated elements and check that correct elements are not edited into errors. As a precision measure analogous to GUARDIAN's false discovery rate, which is reported as often below 20%, we report the fraction of flagged elements that were in fact correct.

The fourth experiment tests transfer. The forecaster is trained on one Caltrans district and tested on another. The Incident Gap papers suggest the forecaster itself will degrade across regions, so the question is whether its calibrated uncertainty widens enough to keep false alarms bounded. We also report the unverifiable rate at each site. A verifier that abstains honestly on sparse sensors is more useful than one that guesses.