acceptodds
Withdrawn from ICLR 2027

MUSE: Multi-Granular Urban Semantic Environment Modeling for Road-Network-Constrained Trajectory Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Human mobility trajectory generation synthesizes spatiotemporal traces that reflect real travel patterns, supporting urban planning, transportation analysis, and mobility research. Existing road-network-constrained generators primarily learn transitions between road segments from historical trajectories, leaving the urban environmental knowledge associated with these roads insufficiently explored. Points of interest (POIs) provide rich descriptions of urban functions, but turning heterogeneous, multi-scale POI information into road-aligned representations for generation remains challenging. To this end, we introduce **MUSE**, a Multi-granular Urban Semantic Environment modeling framework for integrating road-network topology with road-aligned environmental knowledge. It organizes open POI data into three complementary granularities: aggregate functional distributions (Macro), nearby POI instances (Local), and representative functional anchors (Landmark). A Road Query selectively aggregates instances and fuses these sources while preserving POI categories and source identities. Road and Environment autoregressive branches then learn from the same next-road targets, with gated environment-to-road interaction incorporating environmental history into road generation. Experiments on Beijing, Chengdu, and Xi'an show that **MUSE** improves the spatial and temporal distributional fidelity of generated trajectories, demonstrating the value of extracting road-aligned, multi-granular environmental knowledge and learning it jointly with road-network structure.

Primary area:
generative models
Keywords:
Trajectory Generation, Human Mobility Modeling, Autoregressive Modeling, Multi-granular Representation Learning, POI Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MXB6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper frames its task as trajectory generation, although it is closer to routing, as both the origin and destination are specified in advance. Its central argument is that prior work has not made effective use of POI data and therefore has not fully captured the urban context surrounding road segments. The authors aim to better leverage this information to improve the quality of generated trajectories. Specifically, they propose a new method for processing POI data and feed the resulting information into a Siamese Transformer-based architecture to incorporate the urban context along the route into trajectory generation. The reported results show that, given an origin–destination pair, the generated trajectories more closely resemble real-world trajectories in terms of several trajectory-level statistics.

Critical strengths
  1. The proposed POI preprocessing method helps make more effective use of the available data, and incorporating the surrounding urban context into road-segment representations is a sensible design.
  2. The experimental results show clear improvements over comparable baselines, and the ablation studies exhibit consistent overall trends across all three cities.
Critical weaknesses
  1. The paper’s contribution is confined to the narrowly defined task of representing road segments within trajectories, which substantially limits its broader applicability.
  2. The model architecture follows a fairly straightforward design and offers limited methodological novelty. It provides few new insights or design principles that readers could draw on beyond this particular task.
  3. The POI preprocessing pipeline does involve some deliberate design choices. However, it largely amounts to combining a range of existing, straightforward techniques rather than introducing a particularly novel or insightful approach to extracting urban semantics.
  4. The model relies on multiple conditioning inputs, including the origin and destination. It remains unclear whether the reported performance would hold without these additional constraints, which further limits the approach’s applicability. As noted in my summary, specifying both endpoints effectively narrows the task from general trajectory generation to routing between fixed locations.
  5. The paper’s writing and problem framing need further refinement. In particular, the challenges presented in the introduction appear to describe necessary processing steps rather than substantive research challenges.
  6. From the perspective of origin–destination pairs, a strong relationship between trip intention and urban functional context is reasonable. However, once the origin and destination are fixed, route choice is often driven primarily by shortest-path, as reflected by Dijkstra in the paper, with other variations often arising from dynamic traffic conditions and user-equilibrium effects. Urban context may still be relevant, but its role in routing appears more limited and less fundamental than the paper’s motivation suggests.
Questions for authors
  1. Some of the baselines are autoregressive models. How are these models conditioned on the destination? If they are not given the destination, the comparison would be unfair, as the proposed model has access to additional input information that these baselines do not receive.
  2. Do the gains from the environmental branch come from its explicit prediction objective or simply from the additional model capacity? The full model differs from the “w/o Env Branch” variant not only in the inclusion of LenvL_{\mathrm{env}}, but also in the presence of a separate Transformer. This ablation therefore does not disentangle the effect of the prediction objective from that of the increased model capacity.
6 comments
AuthorsOpenReview ↗

Response to Reviewer MXB6 (1/4) — Contributions and methodological novelty

Thank you for recognizing the consistent improvements across all three cities. We would like to clarify the environmental modeling contribution underlying these results.

W1. MUSE contributes an independent environmental modeling framework

Road-segment representations are a foundation of road-network trajectory modeling. MUSE builds on this foundation by modeling Road and Environment separately, from token representation to sequence dynamics, and integrating them at the sequence hidden-state level (Sections 3.2–3.5). Environmental information is not incorporated into the initial Road token. The Road tokenizer represents road attributes, topology, and regional context; the Environment tokenizer learns a separate vocabulary aligned to the same road indices.

This environmental representation preserves both statistical and instance-level information. Macro describes the aggregate composition of nearby POI categories. Instance representations retain facility semantics and road-relative geometry. We further separate instances into Local and Landmark sources to distinguish the contributions of nearby facilities and representative functional anchors. Road Query first aggregates instances within each instance source, then fuses Macro, Local, and Landmark through source attention. This hierarchy provides a road-aligned environmental representation while preserving the distinctions among the three sources during aggregation.

The two token sequences enter separate causal Transformers. The Environment branch learns the encountered environmental history and receives its own next-road prediction supervision. Its hidden states interact with the Road hidden states through a learned gate, and the fused states support road and time prediction. Environmental representation learning, alignment and fusion of data and representations, independent sequence modeling, and engineering design choices such as predictive supervision are all core contributions of MUSE. Describing the work only as road-segment representation leaves out this complete environmental modeling framework.

The broader insight is to make unordered, heterogeneous environmental observations an independently learned predictive view aligned with the units of a structured sequence. This is a general sequence-modeling perspective, instantiated here with roads and POIs. Across three separately trained city models, MUSE improves all four distributional metrics over the road-only variant, with Duration JSD reductions of 32.2%–67.6%. These results establish the value of explicit environmental modeling for trajectory generation and provide a concrete basis for incorporating environmental context into other spatiotemporal modeling tasks.

W2. Methodological novelty and reusable design principles

MUSE contributes new insight at both the representation and sequence-modeling levels: preserve complementary environmental granularities, learn environmental history as its own predictive sequence, and use supervised interaction between the structural and environmental views to generate movement.

The ablations substantiate these design principles. With both branches retained, the full multi-granular representation improves all four JSDs over Macro alone in every city, reducing Duration JSD by 11.0%–32.9%. This establishes the benefit of jointly representing statistical context and individual facilities.

With the same multi-granular environmental inputs, the full model improves all four JSDs over the single-branch feature-concatenation variant, reducing Duration JSD by 19.7%–52.5% (Table 3). This demonstrates the effectiveness of the combined design of independent sequence modeling, hidden-state interaction, and environmental supervision.

The new experiments in Part 4 further isolate the auxiliary objective and compare environmental and road inputs with identical architecture and parameter count. They show that both explicit environmental supervision and meaningful environmental content contribute to the gains. Thus, the representation design, sequence architecture, and training choices yield concrete, experimentally supported guidance that readers can use when integrating heterogeneous context into sequence models. This contribution directly addresses the questions of methodological novelty and informativeness.

AuthorsOpenReview ↗

Response to Reviewer MXB6 (2/4) — Urban semantics, OD, and environmental relevance

W3. The POI design implements a coherent insight

Our insight is that urban semantics must be represented at complementary granularities and learned according to their relevance to movement. Similar category densities can conceal different individual facilities and different relationships between those facilities and a road. Local facilities and representative landmarks also provide different information. A useful environmental representation must retain these distinctions and learn how they matter for each trajectory unit.

The hierarchy follows this reasoning. Instance attention learns relevance within Local and Landmark; source attention learns the contribution of Macro, Local, and Landmark to each road. Candidate construction supplies observations, while the learned tokenizer and trajectory objectives determine how those observations become useful environmental representations. The resulting tokens are then modeled as an independent sequence with explicit predictive supervision.

The source ablations support this insight. Adding Local to Macro improves all four metrics in every city. Adding Landmark to Macro+Local further improves Distance, Radius, and Duration in every city. The full three-source model improves all four metrics over Macro alone. Together with the sequence and supervision ablations, these results validate the effectiveness of the proposed environmental representation and its use in generation.

Existing retrieval, attention, and Transformer operations are the means by which we turn this insight into measurable gains. The use of established operations does not justify reducing the work to those operations or dismissing the contributions of the modeling formulation, the distinctions it preserves, and the supporting experimental evidence. MUSE offers a coherent and reusable approach to learning trajectory-relevant urban semantics, extending well beyond a collection of preprocessing operations.

W4. Fixed OD still leaves a substantive trajectory generation problem

OD specifies the request, not the trajectory. MUSE learns p(trajectory | O, D, t₁): a distribution over timestamped road sequences given origin, destination, and departure time. It must generate intermediate roads, corridors, detours, transition times, and trip durations.

Even with identical endpoints and departure time, trips vary with intermediate activities, preferences, and circumstances. Departure time also changes route and speed distributions. Shortest-path routing optimizes a chosen path cost; trajectory generation learns the distribution and spatiotemporal diversity of observed movement. OD does not reduce this task to shortest-path routing.

OD is part of an established task formulation, not a methodological advantage claimed by MUSE. Autoregressive TS-TrajGen and HOSER and diffusion-based DiffTraj and Cardiff also use OD. Our contribution is the environmental modeling framework and its gains within this task. This setting serves demand-conditioned mobility simulation: a demand model supplies requests, and a generator produces the corresponding trajectories. Applicability should be assessed in this established setting.

W6. Fixed OD does not eliminate the role of the urban environment

Shortest paths and traffic are important influences on travel, but they do not fully describe urban movement. A taxi may change its route because a passenger changes the destination, a road is temporarily closed, or congestion builds during peak hours. A private traveler may avoid busy school or hospital access roads, while a shared ride may involve a pickup outside the most direct corridor. The final recorded endpoints leave these intermediate decisions and their temporal effects unresolved.

These examples illustrate why route choice reflects interacting causes at several levels, including trip activities, access to facilities, preferences, traffic, and network constraints. Urban functions shape where activities occur and how travelers access them. Exploring the causal mechanisms behind urban mobility, including these environmental influences, is a central motivation for MUSE. We start from road-network sequences and explicitly model the surrounding environment to capture information about movement that road topology and endpoints alone do not provide.

The claim that environmental context becomes secondary once OD is fixed therefore needs to be tested empirically. Our experiments directly address it. On the same OD pairs, MUSE reduces Distance JSD by 66.1%–77.8% and regional visitation JSD by 78.0%–81.5% relative to Dijkstra. This establishes a substantial gap between shortest-path outputs and observed mobility distributions. The controls in Part 4 isolate the environmental contribution within the learned generator.

AuthorsOpenReview ↗

Response to Reviewer MXB6 (3/4) — Baseline fairness and research challenges

Q1. Destination conditioning and faithful baseline reproduction

We adapt each released baseline implementation to the city datasets while retaining its published network architecture, training objective, complete native conditioning inputs, and recommended training and sampling settings. Learned baselines are retrained separately for each city on the same 80/10/10 splits. We do not modify their networks to force a uniform conditioning interface. The original mechanisms are as follows.

MethodOrigin and destination interfaceOther inputs or priors
TS-TrajGenO initializes the prefix, D enters cost estimationDeparture time, A* heuristic, regional guidance
HOSERO initializes the prefix, D guides the navigatorDeparture time, road/regional encodings, path search
DiffTrajOD grid embeddings condition denoisingDeparture time, reference road count, duration and distance
CardiffOD grid embeddings enter both diffusion stagesDeparture time, trip length/duration, coarse-road latent
MobilityGPTRoad-prefix continuation, no fixed DConnectivity mask, gravity-based training sampling
Seed20-road prefix, no separate OD pairConditional diffusion
MoveSimNo fixed ODPOI-category similarity, time embeddings, distance/history
DijkstraOD specifies path endpointsRoad lengths and topology
MUSEO initializes the prefix, D token enters both streamsDeparture time, regional features, topology, POIs

In particular, TS-TrajGen and HOSER use the destination through cost estimation and navigation, respectively. The autoregressive comparison therefore includes methods that already use D and their original search or structural priors.

HOSER and MUSE receive the same origin road, destination road, and departure time. Under these shared request conditions, MUSE reduces Duration JSD by 64.6%–79.4%. The distributional evaluation uses three MUSE samples and one deterministic HOSER trajectory per request. The MUSE ablations also retain the same request inputs throughout. These comparisons show that the environmental modeling gains cannot be attributed to providing MUSE alone with the destination.

A further evaluation supplies both models with the same observed road and time prefixes and the true next road at each step. With the route controlled, MUSE reduces whole-trip travel-time MAE by 50.4%–67.5% across all 12 city and route-length groups. Thus, its temporal advantage persists when route choice is fixed. We will include the method-specific conditioning interfaces in Appendix H to make the comparison transparent.

W5. Substantive challenges and clearer presentation

We will follow the suggestion to sharpen the Introduction and make the research challenges more explicit. The underlying difficulty is the mismatch between environmental observations and the trajectory sequence to be generated.

Ambiguous spatial and semantic correspondence. POIs have heterogeneous meanings and spatial supports. One facility may relate to several roads, while a road may have many nearby facilities whose relevance differs. Proximity alone does not resolve these associations. The research challenge is to learn which observations characterize a trajectory unit and how to align them into a useful environmental view.

Complementary information at different granularities. Category statistics preserve functional composition but discard individual facilities; instance sets retain detail but introduce redundancy and uneven coverage. Local facilities and landmarks have different representational roles. The challenge is to preserve useful distinctions while learning their relative contributions, rather than lose information through a single uniform summary.

Learning how environmental history affects generation. Environmental observations are static and unordered, whereas movement is a dynamic sequence of road transitions and times. Merely making POIs available does not teach a model how the encountered environment relates to subsequent movement. The challenge is to learn a predictive environmental history and coordinate it with the road history during generation.

These are representation and learning problems that explain the multi-source hierarchy, independent branches, and environmental supervision. We will make this connection explicit in the Introduction, clarify the two token vocabularies and hidden-state fusion around Figure 2, and connect the design rationale to the ablation evidence.

Reviewer_rCfGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies road-network-constrained trajectory generation, proposing MUSE to enhance trajectory generation with environmental information by considering three levels of environmental semantics: local, macro, and landmark. It further designs corresponding encoding and training strategies to incorporate these heterogeneous information. The proposed method achieves state-of-the-art performance on datasets of three cities.

Critical strengths
  1. The paper integrates environmental information into trajectory generation and considers the fusion of information at multiple granularities.

  2. The proposed method achieves strong performance across three cities, and ablation studies are conducted to validate the effectiveness.

Critical weaknesses
  1. The presentation of the paper could be improved. For example, the introduction section uses three separate short paragraphs to explain the three challenges, but lacks an overall overview of the proposed method. Similar issues can also be found in the method section, making the paper somewhat difficult to follow.

  2. Could the authors clarify whether MUSE the first work to incorporate environmental information into trajectory generation? It seems that MUSE benefits from additional information, but also introduces higher requirements for the input data. It would be helpful to discuss this trade-off.

Questions for authors
  1. See weakness.

  2. The paper provides an overall description of the proposed method, but some implementation details are still missing. Releasing the code would be helpful to ensure reproducibility.

  3. Figure 3 and 4 provide qualitative visualizations. While they look visually appealing, the lack of comparisons makes it difficult to draw meaningful conclusions. Authors may consider including baselines for comparison.

  4. The explanation of Figure 5 is not clear enough, making it difficult to understand the results. It will be helpful to provide more details.

5 comments
AuthorsOpenReview ↗

Response to Reviewer rCfG

Thank you for recognizing MUSE's multi-granular modeling, three-city results, and ablations.

W1. Readability of the introduction and methodology

We agree that both sections need a clearer narrative. We will add an introduction overview linking the three challenges to our solutions, organize the methodology around Figure 2, and clarify module purposes, inputs, outputs, and connections. Consistent terminology and transitions will help readers follow the method.

W2 / Q1. Novelty and additional environmental data

To our knowledge, MUSE is the first work to investigate the complex urban mechanisms behind spatiotemporal trajectory generation through explicit environmental sequence modeling, and to demonstrate the effectiveness of this perspective across cities. Environmental information motivates the entire framework.

MoveSim performs generation, but its use of POIs is limited to a location-similarity prior derived from category distributions. It does not learn the encountered environment as a predictive sequence. CORE models multi-granular environments primarily to encode observed trajectories for downstream tasks, including generation. It summarizes road- and grid-level semantics; MUSE preserves individual facilities and aligns distinct sources to roads. MUSE learns how trajectories unfold from an independent environmental view.

MUSE preserves aggregate composition (Macro), nearby facilities (Local), and representative functional anchors (Landmark). Road Query aligns these sources to roads while retaining instance semantics and road-relative geometry. A separate environment vocabulary feeds a causal Transformer with its own next-road prediction objective. Its hidden states interact with the Road branch to generate both roads and times. Environmental representation, alignment, and independent sequence learning form a complete modeling contribution.

Table 3 substantiates this perspective: MUSE improves all four JSDs over the road-only variant in every city, reducing Duration JSD by 32.2%–67.6%. With the same environmental inputs, the full framework reduces Duration JSD by 19.7%–52.5% over feature concatenation. These results establish the value of the environmental view and a concrete basis for investigating urban mobility mechanisms.

POI cleaning, spatial association, and Landmark selection are automated offline, requiring no manual annotation. Generation reuses these resources without LLM calls. We already discuss the temporal mismatch between POI snapshots and trajectories and will clarify this data trade-off.

Q2. Code and data resources

MUSE's source code is available in our repository. Trajectory, road, and region data for Beijing, Chengdu, and Xi'an come from VecCity, whose official repository provides download links and instructions. POIs come from the Amap POI dataset hosted by the Peking University Open Research Data Platform; downloading requires registration and login. Our code repository includes preprocessing, training, and generation instructions to support reproducibility, reuse, and discussion.

Q3. Baseline comparisons in Figures 3 and 4

We have added HOSER to Figure 4 and will include it in the revision. Figure 3 will also include baseline comparisons with matched spatial extents and color scales.

Q4. Central finding of Figure 5

The central finding is that learned environmental information allocation is consistent with intuitive human travel reasoning, offering a preliminary interpretation of trajectory sequences through their environmental context. We observe this at three levels:

  • Across roads (c): Roads learn different Macro–Local–Landmark mixtures. Differences persist among roads with all three sources available, indicating road-specific environmental preferences.
  • Trajectory roles (b): Combined Local–Landmark weights are 60.1%, 56.0%, and 62.3% at origins, en route, and destinations. Endpoints average about 5.2 percentage points more weight on POI instances, while en-route roads relatively favor Macro. This is consistent with specific facilities identifying endpoint activities and aggregate functional composition characterizing the environment traversed. These are static road weights grouped by trajectory role.
  • Individual entities (a): A campus-access road attends to the school (0.68) over a similarly distant gate (0.04); a residential road favors the community (0.72) over a closer pharmacy (0.18). This reflects function-aware selection consistent with how people understand places, beyond distance ranking alone.

These patterns provide a preliminary model-level interpretation of trajectories through environmental context. We will revise Section 4.4 and the caption around this finding.

AuthorsOpenReview ↗

Updated Manuscript and Supplementary Material

We have updated the manuscript PDF and supplementary material to incorporate your suggestions into the additional experiments and manuscript revisions. Thank you again for your constructive feedback. We respectfully ask you to reconsider your assessment in light of these updates.

AuthorsOpenReview ↗

A brief follow-up on our response

Dear Reviewer rCfG, as the discussion period nears its end, we wanted to gently follow up before the discussion closes. Has our response helped resolve your concerns? We would be happy to clarify anything further and would appreciate hearing your updated thoughts.

Reviewer_3hV4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MUSE, a framework for road-network-constrained trajectory generation that outputs connected road segments with timestamps. MUSE aligns multi-granular POI information, including aggregate distributions, nearby instances, and LLM-selected anchors, to individual roads through a Road Query and two-level attention, yielding an Environment token alongside a Road token that combines attributes, topology, and regional context. Two causal Transformers share a next-road prediction target, a gate injects environmental history into road states, and a candidate-conditioned head predicts time increments. The contributions are a mechanism aligning heterogeneous POIs to roads, a dual-view autoregressive framework giving the environment an explicit predictive role, and empirical gains on three cities. Validation uses trajectories for Beijing, Chengdu, and Xi'an with Amap POIs, conditioned on origin, destination, and departure time, and compares against Dijkstra and nine generators. Evaluation uses Jensen-Shannon distances on four macro distributions (Distance, Radius, Duration, LocFreq), with Duration available for only three methods.

Critical strengths

Operationalizing environmental knowledge as a road-aligned representation. Rather than appending POI features to road vectors, the authors use the road attribute query as a Road Query and two-level attention, at the instance level and the source level, to aggregate three heterogeneous sources onto individual roads, with source-type embeddings and an availability indicator for missing sources. The consequence is that POI knowledge enters generation as a road-indexed token on the same footing as road attributes, so the environment component can be added, removed, and ablated independently, and it can in principle be reused by any road-sequence generator.

Critical weaknesses
  1. The attribution of the environmental branch is not established. Both branches predict the same next-road target, making the environment branch a second prediction head with its own road vocabulary, and the ablation that removes it simultaneously eliminates auxiliary supervision, independent parameters, and the gate. The paper reports neither the branch's own accuracy nor shuffled-POI, random-POI, or parameter-matched controls, so the gain may reflect extra capacity or POI acting as a road fingerprint rather than environmental semantics. Since that semantics is the paper's central claim, the mechanism remains unvalidated.
  2. The comparison is incomplete and, for coordinate-space baselines, not fair. The most relevant recent work is missing: Seed (WWW 2025) on road trajectory generation, Cardiff (KDD 2026) on coarse-to-fine cascaded diffusion, HTP (KDD 2026) on LLM-based semantic trajectory generation. At the same time, the coordinate-space baselines are evaluated under a protocol that disadvantages them: their conditioning interface is undocumented, trajectories are resampled by arc length rather than with their native sampling, no map matching is applied. The main evidence is marginal JSDs on four distributions, with no path-level comparison against baselines. The claim of outperforming all baselines on all four metrics therefore does not hold under a common protocol, and the reported margins cannot be attributed to the proposed method rather than to the adaptation.
Questions for authors
  1. The paper neither cites nor compares Seed (WWW 2025, road trajectory generation with sequence and diffusion models), Cardiff (KDD 2026, coarse-to-fine cascaded hybrid diffusion), or HTP (KDD 2026, LLM-based semantic trajectory generation), and CORE is discussed in only one sentence. Could you explain why these are excluded, and either add them as baselines or justify why they are not comparable?
  2. Do all baselines receive the same conditioning information? Please provide a per-method table listing the exact fields supplied (origin segment, destination segment, departure time, trip distance, region, road topology), how each field enters the model, whether the architecture or training objective was modified, and whether the model was retrained.
  3. The coordinate-space baselines perform far worse here than in their original papers, for example, DiffTraj's Distance JSD is 0.7753, which means the two distributions are nearly disjoint. Could you explain this discrepancy? Is it due to the metric definition (sum of point-to-point distances), the arc-length resampling, the absence of map matching, the conditioning mismatch, or an implementation issue? Please also describe normalization and denormalization, unit handling, and any sanity check confirming that the same pipeline applied to real trajectories reproduces the reference distribution.
  4. Temporal evaluation. Is Duration, a single marginal over whole trips, sufficient to evaluate temporal fidelity? The model predicts per-step time increments, yet no per-segment timing error is reported. Could you report, for example, MAE or RMSE of log time increments on real transitions, speed or travel-time distributions by road type and time of day, and arrival-time error? Since Duration and Distance are strongly correlated, please also report duration error conditioned on distance, so that the temporal gain can be separated from the improved route choice.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 3hV4 (1/2): Attribution of environmental gains and new baselines

Thank you for recognizing the value of road-aligned environmental modeling. Your comments helped us more directly test our central claim: meaningful environmental content and its independent predictive role can improve trajectory generation. We added targeted controls, the three recent baselines you identified, and path-level and fixed-route temporal evaluations. Below, we explain how this evidence addresses your main concerns.

W1. Environmental content and supervision contribute beyond capacity

We added four controls with identical trip-conditioning inputs to examine supervision, alignment, content, and capacity.

  • w/o Env supervision sets the auxiliary Environment loss weight to zero, retaining both Transformers, environmental inputs, the fusion gate, and all parameters.
  • Shuffled POI shuffles and reassigns the same environmental representations to different roads.
  • Random POI replaces environmental content with fixed random features.
  • Road+Road retains both branches, the gate, auxiliary supervision, and exactly the same parameter names, shapes, and count. Only the Environment input is replaced with features derived from road attributes, frozen node2vec embeddings, and road adjacency; no POI content or original POI masks are used.

Each cell reports spatial-average JSD / Duration JSD (lower is better). Spatial average is the arithmetic mean of Distance, Radius, and LocFreq JSD. BJ/CD/XA denote Beijing/Chengdu/Xi'an.

SettingBJCDXA
MUSE0.0258/0.04690.0266/0.04390.0268/0.0280
w/o Env supervision0.0268/0.05010.0272/0.04780.0277/0.0371
Shuffled POI0.0291/0.05560.0286/0.06140.0287/0.0581
Random POI0.0288/0.05450.0288/0.05270.0292/0.0534
Road+Road0.0292/0.05720.0284/0.05580.0287/0.0422

Environmental prediction supervision contributes with capacity fixed. Compared with w/o Env supervision, the full model improves all four JSDs in every city, reducing Duration JSD by 6.4%–24.5%. This directly validates the role of the Environment branch's independent prediction objective.

Environmental content contributes with architecture, supervision, and parameter count matched. MUSE improves all four JSDs over Road+Road in every city, reducing Duration JSD by 18.0%–33.6%. Additional Transformer capacity alone therefore cannot explain these improvements.

Correct road–environment alignment also matters. Shuffled POI preserves the same environmental feature collection but worsens all four JSDs in every city; Xi'an Duration JSD rises from 0.0280 to 0.0581. Fixed random features also fail to reproduce the gains from real environmental inputs.

Together, these controls establish contributions from meaningful environmental content, correct road alignment, and predictive supervision beyond extra parameters or arbitrary fixed road identifiers. They add direct evidence about the mechanism to the original branch ablations, addressing the central concern that environmental attribution was not established.

W2 / Q1. Adding recent baselines and extending evaluation to paths

We added the three requested baselines. Each cell reports Distance / Radius / LocFreq JSD, in that order.

MethodBJCDXA
Seed0.0838/0.1139/0.28140.1731/0.1628/0.23500.1161/0.1365/0.2287
Cardiff0.1315/0.1116/0.17310.0809/0.0700/0.08260.0733/0.0512/0.0583
HTP0.0354/0.0149/0.04870.0759/0.0631/0.03800.0587/0.0307/0.0399
MUSE0.0211/0.0129/0.04340.0353/0.0175/0.02710.0313/0.0221/0.0270

MUSE's spatial distributional advantage persists after adding the recent methods you identified. MUSE outperforms all three added baselines in all nine city–spatial-metric combinations, reducing Distance JSD by 40.4%–53.5% relative to HTP. This further validates the original findings against the expanded baseline set.

We also added destination-attainment and road-sequence consistency evaluations. Destination attainment in BJ/CD/XA is 99.357%/99.707%/99.943%, respectively, compared with 100% for HOSER in all three cities. MUSE's Best-of-3 LCSS is 0.817/0.915/0.901, compared with 0.764/0.867/0.860 for HOSER's single deterministic path. LCSS uses exact road-ID matching and normalization by the shorter sequence length; we average the per-request maxima. These results extend the evidence on generation quality from marginal distributions to destination attainment and path consistency.

CORE focuses on trajectory representation learning for downstream tasks, hence its initial exclusion as a trajectory generation baseline. It distills POIs into textual semantics to enrich road-network embeddings; MUSE aggregates multi-granular POIs via Road Query attention and explicitly trains an independent Environment sequence branch for prediction. We will clarify this distinction and, space permitting, further evaluate CORE's trajectory generation capabilities.

AuthorsOpenReview ↗

Response to Reviewer 3hV4 (2/2): Baseline configuration and spatiotemporal fidelity

Q2. Preserving native designs and comparing matched trip requests

Faithful reproduction requires preserving native inputs and structural priors. We retain each baseline's complete original conditioning inputs, architecture, and training objective, without changing its network to impose a common interface. Learned baselines are retrained per city on the same 80/10/10 split. The differences below come from the original methods; preserving these designs retains their capabilities.

Representative interfaces (O/D: origin/destination; No: unused):

MethodRepresentationPOI useOD useOther conditions / priors
DijkstraRoadNoPath endpointsRoad lengths and network topology
MoveSimDiscrete location IDsSimilarity matrix of POI category distributions in RegNetNo fixed ODTime embeddings; distance and historical-transition matrices
TS-TrajGenRoadNoO initializes prefix; D enters cost estimationDeparture time; A* heuristic; regional guidance and road topology
DiffTrajGPS coordinatesNoOD grid embeddings condition denoisingDeparture time; reference road count, duration, and distance
HOSERRoadNoO initializes prefix; D guides navigationDeparture time; road/regional encodings; path search
MobilityGPTRoadNoPrefix continuation; no fixed DConnectivity mask; gravity-based training sampling
SeedRoadNoPrefix conditioning; no separate OD pair20-road prefix; conditional diffusion
CardiffRoad + GPS coordinatesNoOD grid embeddings in both diffusion stagesDeparture time; trip length/duration; coarse-road latents guide GPS generation
MUSERoadRoad Query aggregates multi-granular POIs into an Env sequence with predictive supervisionO initializes prefix; D token in both branchesDeparture time; regional features/node2vec; adjacent-road candidates

Several baselines already use OD and distance/search priors. HOSER and MUSE receive exactly the same origin road, destination road, and departure time. Thus, the comparison preserves native designs and includes a strong baseline with identical request conditions. Road+Road in W1 additionally fixes architecture and parameter count. Together, these comparisons substantiate the environmental modeling contribution.

W2 / Q3. Coordinate preprocessing and DiffTraj

We concatenate road geometries in traversal order, remove duplicate shared-endpoint coordinates, and linearly interpolate to each model's required length. Source sequences are thus already road-aligned; preprocessing is deterministic and reproducible. DiffTraj normalizes lon/lat by city-specific means/stds, denormalizes after generation, and computes geographic metrics in kilometers. Generated coordinates are scored directly without map matching, separately from training-data road alignment.

We favor interpreting the gap through dataset/task differences and model inductive biases. Diffusion excels at diverse coordinate synthesis; this task also requires connected road transitions, which MUSE explicitly enforces during autoregressive generation. Distance sums consecutive displacements, whereas Radius measures spatial extent. DiffTraj can therefore match Radius (0.0229 in Beijing) yet differ substantially in Distance (0.7753). Matching spatial extent does not ensure correct trip lengths; these results support explicit road-transition constraints for this task. We will report the requested real-trajectory sanity check in Appendix H.2.

Q4. Temporal gains persist with routes fixed

The new evaluation directly controls for route choice. Both models receive identical real road/time prefixes and the true next road. We report log(1+Δt) MAE (Δt in seconds) and whole-trip time MAE (seconds), grouped by reference-route length.

CityMethodlog MAE0–3 km3–6 km6–10 km≥10 km
BJMUSE0.265178.37151.75327.01832.58
BJHOSER0.2907213.32400.82784.522057.68
CDMUSE0.344434.8349.5347.9138.81
CDHOSER0.335791.44137.28122.2878.30
XAMUSE0.386036.7955.8270.6063.90
XAHOSER0.388596.41166.38216.25196.38

MUSE reduces whole-trip time MAE by 50.4%–67.5% in all 12 groups. HOSER's single-step error is slightly lower in Chengdu, but MUSE retains its whole-trip advantage in every length group across all three cities. Thus, timing gains persist both for the Duration marginal and for distance-stratified errors on fixed real routes. This directly addresses whether timing gains merely follow from better route choice.

The added controls, recent baselines, and fixed-route timing evaluation directly substantiate environmental attribution, comparison completeness, and temporal modeling.

AuthorsOpenReview ↗

Updated Manuscript and Supplementary Material

We have updated the manuscript PDF and supplementary material to incorporate your suggestions into the additional experiments and manuscript revisions. Thank you again for your constructive feedback. We respectfully ask you to reconsider your assessment in light of these updates.

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

Related papers

Discussion (0)

Sign in to comment.