acceptodds
Under review as a conference paper at ICLR 2027

Toward a Traffic World Model: Learning the Physics of Signalized Roads with Guided Diffusion

Abstract

Traffic engineers describe how vehicles move along a road with physics-based models built from a few equations. These models are compact and interpretable, but each fits one kind of measurement and must be calibrated by optimization before use. We ask whether a generative model trained only on traffic states and the corresponding signal timing can learn the physics of a signalized road, and whether it can serve as a world model that answers questions from partial sensor data. We present DiffLink, a diffusion model over the time-space diagram of a signalized link. The state is a two-channel image of density and flow, the signal is the action, and no traffic equation enters training. On SUMO data, states generated from the signal alone reproduce queue formation and discharge. Their mean arrival rate, delay, and maximum queue length match the test set within 4%, and their fundamental diagram, discharge flow, and queue-tail growth rate match the simulator’s. When the green is shortened, the model’s queue grows by 4.9 m against the simulator’s 5.1 m, at a correlation of 0.93 across the 32 simulations. Because the model is a prior over full states, any sensor becomes a query at sampling time, with no retraining. Guided by two stop-bar detectors, DiffLink estimates the arrival rate as well as a vehicle count does, and it reconstructs the maximum queue length per window with a correlation of 0.74, against 0.51 for deterministic queuing from the same counts. We close with what still separates DiffLink from a full traffic world model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
learned physics simulation, diffusion posterior sampling, action-conditioned generation, traffic flow model, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CM27OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new method based on diffusion to model and query the traffic state. A central idea is to train a single unconditional-with-respect-to-sensors traffic prior and introduce sensor information only at inference through diffusion posterior sampling. The authors evaluate whether the model has learned traffic mechanics through several probes. It gives some interesting on this two-lane signalized link scenario.

Critical strengths
  1. The paper asks a clear and scientifically interesting question. Rather than reporting only generative reconstruction metrics, the paper asks whether a data-driven generative model has actually learned traffic dynamics.
  2. The experiment is extensive. The evaluation goes substantially beyond summary-statistic matching.
  3. The code is provided.
Critical weaknesses
  1. The experiment scenario is too simple. Consider experimenting on a lot of real world large scale maps.
  2. The idea to apply diffusion on traffic state estimation is not novel. The PMA-Diffusion has explored this direction.
  3. The paper argues that the generative model can replace a physics-based traffic model, but does not compare against a strong physics-based model
Questions for authors
  1. How much performance is retained with substantially fewer DDIM steps?
  2. Can you compare against a calibrated CTM or kinematic-wave state estimator under the same detector configurations?
  3. How does counterfactual performance change when demand is provided explicitly rather than inferred from detector guidance?
3 comments
AuthorsOpenReview ↗

New experiments (CTM baseline, fewer steps, explicit demand, real-world test) and novelty vs. PMA-Diffusion

Thank you for the review. We have revised the paper to address your comments, which we believe made it a better paper.

Q2 / W3. Calibrated CTM baseline. Added (Sec. 4.4, Table 3, App. M). The model has a triangular fundamental diagram on the 32 cells of the link, with the signal at the downstream boundary. Its four parameters are fitted on held-out windows (14 m/s, 1.2 veh/s, 0.30 veh/m, 1.0 veh/s). It takes the demand that deterministic queuing derives from each sensor, and it inserts detector densities where that helps on the held-out windows. Queue correlation on the 256 test windows, CTM vs DiffLink (median of 8 samples):

  • four rows 0.85 vs 0.98, stop bar 0.66 vs 0.81, upstream 0.85 vs 0.93, 30 s outage 0.43 vs 0.77
  • mid-link rows 0.77 vs 0.73, 30 s counts 0.66 vs 0.71, arrival rate only 0.72 vs 0.62

DiffLink is ahead with every layout within 56 m of the stop bar, and the two are close on the mid-link rows and counts. The CTM is ahead only with the arrival rate alone. Given the true inflow and initial state it reaches 0.92, so most of its error comes from the unknown timing of arrivals. It runs all 256 windows in 0.1 s on one CPU core. It has no natural way to use probe trajectories, while DiffLink adds them as a second loss (stop bar 0.75 to 0.88, Sec. 4.5).

Q1. Fewer DDIM steps. App. L plots every sensor's estimate against the number of steps. With the guidance step scaled as 0.5 × 1000 / n, 100 steps stay within 0.06 of 1000 for every sensor, at 83 ms against 827 ms per window on one H100.

Q3. Counterfactual with explicit demand. We give DiffLink each window's arrival rate as a number, with no detector row, over four shifts of green. Its queue grows by 0.75 m per second of green removed, against 0.87 m in the simulator and 0.83 m with upstream detectors (Sec. 4.3, App. J). The CTM given the same arrival rate grows by 0.65 m.

W2. Novelty against PMA-Diffusion. PMA-Diffusion is the closest work, and we cite it. Our paper differs in two ways. It models density and flow on a signalized link, where the signal enters as an action, so the model answers counterfactual questions about signal timing and gives queue and delay. It also tests what the prior has learned, with physics probes, a lookup-table control, and cycle lengths outside training.

W1. Simple scenario. We agree, and we added a zero-shot test on a real-world dataset (App. N, Table 15). The SUMO-trained prior runs without retraining on 160 windows from NGSIM Peachtree Street in Atlanta, a real arterial with turn bays, driveways and platoons from upstream signals. With stop-bar detectors, DiffLink's queue correlation is 0.64, against 0.31 for deterministic queuing, 0.46 for the calibrated CTM, and 0.81 in simulation. During a 30 s outage it is 0.69, against 0.34 for the CTM. With probes added it reaches 0.87. So the prior transfers to real-world data without retraining. Networks of links remain future work (Sec. 5).

Thank the authors for the detailed rebuttal. The single experiment on the simple NGSIM dataset is still not convincing. I therefore keep my original score of 2: weak rejection.

AuthorsOpenReview ↗

Follow-up on the real-world test: choice of dataset

Thank you for reading our response. We would like to explain why the test uses NGSIM, because the dataset was not chosen for convenience.

The test needs five things from a dataset:

  • trajectories of every vehicle on the link, to build the true time-space diagram
  • the recorded signal timing, which is the model's action
  • a field of view that reaches the queue tail, not only the intersection area
  • at least 15 minutes of continuous recording, to cover several signal cycles
  • public access

We screened ten datasets against these criteria: FLUID, CitySim, CityArtery, SIT-PET, DRIFT, the Utah LiDAR dataset, pNEUMA, SinD, HDI and NGSIM Peachtree. Nine of them fail at least one. DRIFT, pNEUMA, HDI and the Utah LiDAR dataset have no signal data. FLUID, SIT-PET and SinD do not cover the queue tail. CityArtery is shorter than 15 minutes, and CitySim is not public. NGSIM Peachtree is the only one that supports the test, although its trajectories carry known tracking noise.

Peachtree Street is also not a simple site. It is an urban arterial with turn bays, driveways, and platoons from upstream signals, none of which are in the training simulations. The test covers eight approaches in two periods, and nothing is retrained or tuned on it.

We agree that one corridor does not settle how far the prior transfers, and Sec. 5 says so. Within the public data that exist today, this is the test that can be run.

Reviewer_bLwAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DiffLink learns a signal-conditioned diffusion prior over density–flow time-space fields for a single signalized road link. Detector and probe observations enter as losses during sampling, allowing the same trained prior to support different sensor layouts, state estimation, and short-horizon forecasting. SUMO experiments examine generated traffic statistics, responses to signal changes, and downstream reconstruction. The model improves queue estimation over deterministic queuing and supports several queries without retraining. Layout-specific supervised networks remain more accurate for point estimation, and the paper reports limitations in uncertainty calibration and probe-based estimation.

Critical strengths
  1. One prior supports several traffic queries without retraining. Separating observation losses from the prior lets the same model handle detector layouts, probe observations, and forecasting. With stop-bar detectors, queue correlation improves from 0.51 for deterministic queuing to 0.74; the forecasting query reaches 0.89. These results establish useful behavior beyond visually plausible generation. The contribution is the reusable probabilistic traffic representation and query interface, even though the diffusion and guidance components are established methods.

  2. The evaluation tests what the prior has learned and makes failure cases visible. The lookup control shows that matching traffic statistics alone does not distinguish generation from retrieval. The signal intervention and downstream queries then provide more discriminating tests: with upstream observations anchoring demand, the predicted mean queue increase is 4.9 m against 5.1 m in SUMO, with correlation 0.93 across 32 simulations. The authors also report the slope error, stronger supervised point estimators, under-coverage, and probe-induced bias. This combination of controls and negative results makes the paper informative and supports my weak-accept recommendation within its single-link simulation setting.

Critical weaknesses
  1. The comparative benefit of reuse is not fully established. Table 2 favors layout-specific supervised U-Nets for queue estimation: 0.97 versus 0.90 with four detectors, 0.96 versus 0.83 with upstream detectors, and 0.78 versus 0.74 with stop-bar detectors. A single mask-conditioned supervised network could also reuse its parameters across layouts, but is not evaluated. The paper demonstrates that DiffLink can answer multiple queries; it does not yet establish a better accuracy–flexibility trade-off than a shared supervised baseline. This limits the comparative claim, rather than invalidating the demonstrated interface.

  2. The action-response and probabilistic interpretations need narrower boundaries. In Section 4.3, the response depends on an upstream demand anchor: without it the mean queue change is only 0.3 m, and the fitted response slope is 1.05 versus 0.60 m per second of green removed. Training demand also depends on the green split, so the experiment supports a conditional response more clearly than autonomous causal dynamics. Section 4.4 reports 66% queue coverage for a nominal 80% interval, and Section 4.5 reports 16–23% downward arrival-rate bias after adding probes. The samples should therefore be treated as approximate guided reconstructions wit

Questions for authors
  1. Section 4.4 states that the supervised model cannot answer the questions in Sections 4.2 and 4.3, whereas Section 4.3 reports its detector-conditioned response to a shorter green. Does the claimed distinction concern generation without observations, rather than the anchored counterfactual? Please reconcile these statements and identify the capability supported by the comparison.

  2. Do you have results for a single mask-conditioned supervised model across layouts, including an unseen layout? If not, can you narrow the comparative claim to the demonstrated reuse of DiffLink rather than an established advantage over shared supervised models? This distinction determines how much methodological significance I assign to the interface.

  3. How much do the guided uncertainty and probe results depend on the evaluation choices? Please clarify whether the coverage conclusion persists when windows from the same simulation are resampled together, and report queue and delay errors in physical units with and without probes if available. These answers would help determine whether the probabilistic and multi-sensor benefits survive beyond correlation-based summaries.

1 comment
AuthorsOpenReview ↗

New experiments on shared supervised baselines, unseen layouts, and coverage by simulation

Thank you for the detailed review. We revised the paper based on your comments, and we believe they have improved its quality. Every point below is in the revised PDF in blue.

W1 / Q2. One mask-conditioned supervised network, and an unseen layout. We trained two shared U-Nets that take the detector mask as input, with 1 to 6 rows drawn per window from all 32 rows or from the 8 rows within 56 m of the stop bar. Queue correlation on the 256 test windows (Table 3):

SensorDiffLinkone layoutshared, allshared, ≤56 mCTM
four rows0.980.970.970.970.85
stop bar0.810.780.820.830.66
upstream0.930.960.960.950.85
mid-link, 120 and 128 m0.73–0.530.36*0.77
30 s outage0.770.47*0.46*–0.43
30 s counts0.710.45*, retrained 0.800.50*–0.66

* outside the network's training. DiffLink is the median of 8 samples (one sample gives 0.98, 0.75, 0.92, 0.69, 0.71, 0.65). CTM is the calibrated cell transmission model added for Reviewer CM27.

On layouts they were trained on, the supervised networks match DiffLink on the queue and beat it on the arrival rate and the delay (App. I). DiffLink estimates better when the sensor changes or fails after training. Sec. 4.4 now says this, and the abstract and the conclusion claim the advantage only for sensor changes.

W2. The response depends on demand. We agree, because training demand scales with the green split, and Sec. 4.3 now says so. We repeated the counterfactual for four shifts, from 8.5 s less to 3.9 s more green per window, with the same seeds and demands. Over 128 pairs, the queue grows by 0.87 m per second of green removed in the simulator, 0.83 m for DiffLink with upstream detectors (correlation 0.95), and 0.77 m for the supervised network. Given only the arrival rate as a number, DiffLink still gives 0.75 m. So the mean response comes from the prior, and the detectors add the differences between simulations (correlation 0.95, against 0.34 without them). These slopes replace the single-shift fit of 1.05 vs 0.60 (Fig. 3b, App. J).

W2 / Q3. Coverage. We now report coverage on all 256 windows, resampling whole simulations (App. D). One conformal factor per quantity, fitted on 32 held-out simulations, gives 79 to 84% coverage for every variant. Before the correction, the 80% intervals cover 71% of queues with stop-bar detectors and 54% with probes.

Probe bias. The revision uses one detector weight of 3.2, chosen by the rule of Sec. 3.3, which anchors the arrivals better than the 0.2 the submission used with probes. The arrival bias with probes falls from 16 to 23% to about 10% in one sample, and to 2% in the median of 8. With stop-bar detectors, probes now raise the queue correlation from 0.75 to 0.88 and lower the queue MAE from 6.2 to 4.0 m and the delay MAE from 8.6 to 5.6 s. With four rows, already at 0.98, they change little (Table 2, Sec. 4.5, App. K).

Q1. Sec. 4.4 against Sec. 4.3. We revised that sentence. The supervised network cannot generate states from the signal alone (Sec. 4.2). In the counterfactual it needs detector rows, and with them it does as well as DiffLink. DiffLink also answers from the arrival rate alone.

Q3. Resampling by simulation. The conclusions hold. App. I and App. K give intervals by simulation for the rows of Table 2. They are slightly wider, and no comparison changes.

Reviewer_5pL3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DiffLink, a diffusion-based generative framework for traffic state estimation and forecasting on signalized roads. It learns a traffic prior from simulation data and incorporates heterogeneous sensor observations through inference-time guidance, enabling a shared model to support multiple estimation tasks without retraining. Experiments demonstrate promising results in recovering traffic dynamics and responding to changes in signal timing, although generalization beyond the training distribution and practical deployment remain insufficiently established.

Critical strengths
  1. A unified generative formulation. DiffLink formulates traffic state modeling as a conditional generation problem, providing a common framework for traffic state estimation and forecasting rather than designing a separate model for each task.

  2. Physics-informed evaluation. The paper evaluates not only statistical similarity but also traffic dynamics and responses to signal changes, offering multiple perspectives on the behavior of the learned traffic prior.

Critical weaknesses
  1. The paper claims that DiffLink improves sensor adaptability and reduces the calibration burden of existing traffic models, but these benefits are not adequately demonstrated. Although inference-time guidance allows a shared generative prior to accommodate different observation types, observation models and guidance weights may still require redesign or retuning. Systematic evaluation across unseen sensor configurations, with quantitative comparisons of adaptation effort and cost against existing methods, is needed to establish whether DiffLink genuinely reduces calibration burden or merely shifts it to inference-time configuration.

  2. The evaluation is restricted to a single simulated signalized road link and short temporal windows, limiting the evidence for generalization to realistic traffic systems. Moreover, task-specific supervised models achieve competitive or better performance on individual estimation tasks, leaving the benefits of DiffLink's cross-task flexibility unclear. The paper does not sufficiently establish that this flexibility offers meaningful advantages in accuracy, computational efficiency, or adaptation cost, particularly given the reliance on simulation data and the relatively expensive diffusion-based inference.

Questions for authors

1.The evaluation is limited to a single 256m two-lane link, 96s windows, and simulation data. Although the authors frame the work as a “first step,” the introduction’s world model framing remains insufficiently operationalized. Furthermore, the OOD signal-cycle experiments reveal mismatches in the arrival-rate distributions under 60s and 140s cycles, with reported KS statistics of 0.23 and 0.31, respectively, exceeding the stated 5% critical value of 0.12. While local traffic dynamics appear to transfer, these results suggest limitations in extrapolating demand distributions beyond the training range. How this method generalizes to unseen signal timings, and what assumptions about training coverage would be required for deployment?

2.DiffLink accommodates different sensor observations through inference-time guidance without retraining the generative prior. However, it remains unclear when this flexibility provides a tangible advantage over existing methods. What are the intended deployment scenarios, and can the authors quantify the resulting benefits in terms of adaptation effort, computational cost, or robustness to changes in sensor configurations?

1 comment
AuthorsOpenReview ↗

New experiment on calibration burden, unseen sensors and timings, and a real-world zero-shot test

Thank you for the review. Your comments led to several changes in the revision, and we think the paper is stronger for them.

Does DiffLink reduce the calibration burden, or move it to the guidance weights? (W1, Q2) A new sensor needs only a loss and one weight, and we now measure that cost. We choose the weight on 128 held-out windows by one rule (highest queue correlation with the mean arrival rate within 5%), never on test data. A sweep takes 8 to 16 GPU-minutes on one H100, and the four detector layouts we tested share one weight. A supervised network takes 0.9 to 1.7 GPU-hours to train and needs training data that contain the new sensor (App. I).

Unseen sensors (W1). Table 3 compares DiffLink with supervised networks shared across layouts on three changes after training. Queue correlation, DiffLink (median of 8 samples) vs a shared network without retraining:

  • unseen layout: 0.73 vs 0.36
  • 30 s detector outage: 0.77 vs 0.46
  • 30 s counts with no occupancy: 0.71 vs 0.50 (a network retrained on counts reaches 0.80)

On layouts it was trained for, the supervised network is as good or better. The use case is a sensor that changes or fails after training, when retraining is not possible. Two sensors can also be combined without retraining. Probes added to stop-bar detectors raise the queue correlation from 0.75 to 0.88 (Sec. 4.5). The calibrated cell transmission model (App. M) is close to DiffLink at the mid-link layout (0.77 vs 0.73) and worse during the outage (0.43 vs 0.77).

Inference cost (W2). 100 DDIM steps stay within 0.06 of 1000 for every sensor, at 83 ms against 827 ms per window on one H100 (App. L). The median of 8 such samples costs about 0.7 s per window.

Unseen signal timings (Q1). At cycle lengths outside training, the local physics and the queue estimate transfer. With four detectors the queue correlation is 0.90 and 0.95 at 60 and 140 s, against 0.47 and 0.66 for deterministic queuing. The KS values you point to concern demand generated from the signal alone, which does not extrapolate. With a sensor that supplies the demand, the estimate holds. Sec. 5 states this.

Single simulated link (W2, Q1). We agree this was the main limitation, so we added a zero-shot test on a real-world dataset, NGSIM (App. N, Table 15). NGSIM recorded every vehicle on Peachtree Street in Atlanta for two periods of 15 minutes. We built 160 windows of 96 s from eight approaches and their recorded signals. The prior, the weights and the median of 8 samples are those of Table 3, with no retraining or tuning. Queue correlation on the six signalized approaches, DiffLink vs deterministic queuing, the calibrated CTM and the supervised network trained for the stop bar:

  • stop-bar detectors: 0.64 vs 0.31, 0.46 and 0.72
  • 30 s detector outage: 0.69 vs 0.25, 0.34 and 0.64
  • 30 s counts: 0.34 vs 0.31, 0.46 and −0.02

With the stop bar and with the outage, DiffLink is above deterministic queuing and the CTM, and the paired intervals exclude zero. With the stop bar alone, the interval of the difference from the supervised network includes zero. Probes added to the stop bar raise DiffLink to 0.87. So the prior trained only on simulation transfers to real-world data without retraining.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes in the revised PDF

We thank the reviewers for their careful reviews. Based on their comments, we added several experiments and a test on a real-world dataset. We believe the changes improve the paper's quality. New and changed text is marked in blue.

Here is a summary of the main changes:

  1. New baselines (Sec. 4.4, Table 3, App. I, M). Supervised networks shared across detector layouts, and a calibrated cell transmission model (CTM).
  2. Sensor change after training (Sec. 4.4, Table 3). An unseen detector layout, a detector outage, and aggregated counts. In all three, DiffLink estimates the queue better than supervised networks that are not retrained.
  3. New detector weight, chosen on held-out windows (Sec. 3.3, App. I). Based on our selection rule, the revision uses one detector weight for every layout, with or without probes.
  4. Median of 8 samples as DiffLink's point estimate in Table 3, at lower cost than one 1000-step sample.
  5. Fusing detectors and probes (Sec. 4.5). Probes raise the stop-bar queue correlation with no retraining.
  6. Counterfactual over four shifts of green, also with demand given as a number (Sec. 4.3, Fig. 3, App. J).
  7. Calibration of the sample spread on all test windows (App. D), errors in physical units (App. K), and fewer sampling steps (App. L).
  8. A zero-shot test on a real-world dataset (App. N, Table 15). On the NGSIM dataset, with no retraining, DiffLink's queue correlation with stop-bar detectors is above those of deterministic queuing and the calibrated CTM, and it is higher still with probes added.

The supplementary code has a script for every new result.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.