acceptodds
Under review as a conference paper at ICLR 2027

Looking Ahead and Back: Anticipatory Generative Modeling for Road-Network Trajectories

Abstract

Synthetic trajectory generation underpins urban applications such as traffic management and urban planning. In the real world, vehicles are constrained by the physical boundaries of the road network and make sequential routing decisions at intersections navigating toward their destinations. However, state-of-the-art methods primarily focus on simulating location sequences, often producing trajectories that drift off the road network and deviate substantially from the realistic. In this paper, we advocate a fundamental paradigm shift: moving beyond mere location simulation toward route planning on the road network. It is particularly challenging at intersections, where multiple outgoing transitions may appear equally plausible given the observed history yet lead to markedly different downstream routes. To address this challenge, we revisit stochastic dynamics on graphs and introduce a new graph-based generative paradigm—Anticipatory Generation—which resolves such decision ambiguity by constructing a short-horizon trajectory continuation (anticipation) before committing to an action prediction. Specifically, we present an anticipatory Foresight model that generates plausible trajectories through progressive action prediction on the road network. In Foresight, we formulate a probabilistic model to incorporate anticipation, construct fiber bundles for the transitions reconciling road topologies, and embed recurrent anticipation updates into the generation process. Extensive experiments on real-world datasets demonstrate that Foresight wins state-of-the-art baselines across diverse metrics. Codes are available at https://....

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Trajectory Generation, Generative Model, Graph Planning

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_w6MzOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to design a method that goes beyond the autoregressive scheme and enables the trajectories on the generated road network to strictly abide by the transfer laws on the road network. But in fact, this is essentially what the classic route choice modeling research in traditional transportation aims to solve, so it is not a brand-new research direction. The method design is ingeniously designed, which can ensure that the generated trajectory is correct in considering the overall general direction while guaranteeing the randomness of the generated trajectory. A relatively good trade-off has been achieved between constraint modeling and distribution modeling. The experimental results reported in this paper also prove that the method does indeed outperform the existing SOTA baselines in some statistical indicators.

Critical strengths
  1. The method is designed well. By modeling a distribution over future movements, it estimates the trajectory’s overall direction and uses this information to guide decisions at the current intersection. Combined with a data-driven learning approach, this design produces trajectories that reflect the transition patterns observed in the training data while maintaining a coherent overall direction. The method also incorporates strict constraints to ensure that all generated transitions are valid on the road network.
  2. The paper presents a rigorous and well-developed formalization, demonstrating technical sophistication and a deep understanding of the problem, particularly its local decision-making structure. The presentation is thorough while remaining clear and readable.
  3. The experimental design is comprehensive and rigorous. Even setting aside the reported results, the evaluation protocol itself provides a sound basis for assessing the model’s performance across multiple dimensions.
Critical weaknesses
  1. The problem addressed in this paper differs substantially from conventional trajectory generation. With both the origin and destination specified, the task effectively reduces to route choice, where the central concern is minimizing travel costs while accounting for user-equilibrium effects. This may explain why even Dijkstra’s algorithm, which does not learn from data, performs reasonably well relative to the other baselines across the reported metrics. The paper should thoroughly discuss its relationship to route-choice research and clearly position its contribution within that scope, rather than presenting it as a contribution to general trajectory generation. Even with this substantial reframing, the claimed paradigm shift in trajectory generation remains unsubstantiated.
  2. Although the method incorporates some clever design choices, the overall framework lacks elegance, and it is difficult to extract transferable insights that could inform research beyond this specific task. From a technical perspective, the contribution falls short of the level of generality expected at ICLR. The design is highly task-specific, and its broader applicability is limited.
  3. Several aspects of the experimental results warrant closer scrutiny. For example, in the main-text ablation study, removing certain components leaves the backbone and the remaining proposed modules unable to outperform even Dijkstra’s algorithm, despite the latter requiring no learning from data. In addition, the standard deviations reported in Tables 14–18 of the appendix show unexpectedly similar values in some cases and unexplained inconsistencies in others. These patterns raise concerns about the reliability of the reported results and require further clarification.
Questions for authors
  1. How do autoregressive methods, such as the most basic Markov, incorporate origin and destination constraints to ensure fair comparison?
  2. How does the ww parameter in Formula 11 affect the performance in the subsequent experimental table?
8 comments

We thank the reviewer for the careful assessment and for recognizing the strengths of FORESIGHT's methodological design, formalization, and evaluation protocol. We appreciate the opportunity to clarify the relationship to classical route-choice modeling, the transferable principle behind Anticipatory Generative Modeling (AGM), the interpretation of the ablation results, and the baseline and parameter settings. We also address the privacy concerns raised in the review.

1. Relationship to Classical Route-Choice Modeling (CW1)

1.1 Problem scope: generative route choice under OD conditions

We agree that destination-conditioned trajectory generation on road networks is closely related to classical route-choice modeling. Given an origin and destination (OD), a generated trajectory can indeed be interpreted as a sampled route among feasible alternatives. We do not claim that OD-constrained route choice is a new problem. Our contribution concerns the generative mechanism used to learn and sample complete trajectories from observed mobility data.

It is also useful to distinguish individual route choice from user-equilibrium traffic assignment. Individual route-choice models describe how travelers select among available routes; user-equilibrium models additionally study network-level interactions in which aggregate flows affect route costs. FORESIGHT addresses the former, at the level of observed individual trajectories. It does not model congestion-dependent network equilibrium or seek to establish user-equilibrium conditions.

Classical route-choice research includes deterministic shortest-path methods, stochastic utility-based approaches, and data-driven models. We therefore do not suggest that classical methods are exclusively deterministic or that learning heterogeneous route preferences is unique to FORESIGHT. Our objective is to learn a conditional distribution over complete, road-network-feasible trajectories,

Pθ(X0:τ∣o,g,G),P_\theta(X_{0:\tau}\mid o, g, G),

where (GG) denotes the road network and (τ\tau) is the variable termination time. The target distribution concerns complete trajectories from (oo) to (gg), while incomplete generations are accounted for separately as failures in evaluation. The distribution should reflect multiple plausible routes, including observed non-shortest alternatives and the statistical structure of route choices under the same OD conditions. In contrast, a deterministic shortest-path algorithm returns a cost-minimizing route for its specified edge costs; it does not itself model a distribution learned from observed route choices.

Thus, we position FORESIGHT as a generative modeling approach to destination-conditioned route choice, rather than presenting route selection under OD constraints as a newly introduced task. The distinctive contribution lies in how the current generative decision is constructed.

1.2 Explicit future anticipation in current transition selection

A standard sequential generator can directly parameterize a next-state distribution from the realized prefix (ctc_t) and destination (gg):

P(St+1∣ct,g).P(S_{t+1}\mid c_t,g).

We do not claim that this factorization is inadequate or that autoregressive models cannot, in principle, represent long-range dependencies. AGM retains sequential factorization but introduces a different computational organization of each transition decision.

FORESIGHT first constructs a provisional representation of future evolution,

R^t=Fθ(ct,g),\hat R_t=\mathcal F_\theta(c_t,g),

and then uses this anticipation to form a distribution over the topologically legal next states,

$q_t(s')=\mathcal T_\phi(s'\mid c_t,\hat R_t,\mathcal U_t), \qquad s'\in\mathcal U_t.$

Here, (\mathcal U_t) is the feasible candidate set derived from the current road-network state. Following the selected transition, the realized prefix is updated and anticipation is recomputed:

$ c_t\ \longrightarrow\ \hat R_t\ \longrightarrow\ q_t\ \longrightarrow S_{t+1}\ \longrightarrow\ c_{t+1}\ \longrightarrow\ \hat R_{t+1}. $

The anticipated continuation is neither a known future trajectory nor a path that the model commits to following. It is learned predictive information used to inform the probability of the current legal transition. This mechanism differs from a model that directly scores the next transition using only a history-derived representation, without asserting that all existing route-choice models ignore downstream information.

FORESIGHT also uses candidate-conditioned re-anticipation during learning. For each legal movement candidate (s′s'), it forms an updated hypothetical prefix and predicts a continuation,

$ \hat R_{t+1}^{(s')}=\mathcal F_\theta(c_t\oplus s',g). $

The corresponding continuation prediction errors are weighted by the current transition probabilities in the learning objective. This connects the learned next-transition distribution to downstream predictive behavior. These candidate-conditioned continuations are not claims about true counterfactual outcomes or causal effects; nor are they required to be supplied as inputs to the Transition Network's base forward scores.

Accordingly, the contribution is not road-network feasibility alone, but the integration of explicit anticipation, feasible probabilistic selection, candidate-conditioned continuation learning, and closed-loop re-anticipation. We will clarify this positioning in the Introduction and Related Work, including its relationship to stochastic route-choice methods that already account for downstream route information.

2. Generality and Transferability of AGM (CW2)

We understand the concern that FORESIGHT's architecture is specialized to discrete road networks. We distinguish the general modeling principle of AGM from the domain-specific components of its FORESIGHT realization.

2.1 A transferable generative abstraction

Consider a sequential generative problem with a realized context (ctc_t), conditioning information (zz), a feasible decision set (At\mathcal A_t), and a context-update operator (U\mathcal U). An AGM-style construction can be expressed as

$ \hat R_t=\mathcal F_\theta(c_t,z), $

$ q_t(a)=\mathcal T_\phi(a\mid c_t,\hat R_t,\mathcal A_t), \qquad a\in\mathcal A_t, $

$ a_t\sim q_t,\qquad c_{t+1}=\mathcal U(c_t,a_t). $

The key modeling choice is to explicitly predict information about subsequent evolution before constructing the current feasible-decision distribution, and to rebuild that prediction after the realized decision is appended to the context. This establishes a separation between (i) future anticipation, (ii) current decision distribution, and (iii) state update and re-anticipation.

In FORESIGHT, (z=gz=g), the decision set is derived from outgoing road-network connections (with STOP allowed only at the destination), and a finite-horizon decoder realizes the anticipation operator. These choices are specific to road-network trajectories. The abstraction itself does not require a road graph: another application could supply a different feasible decision set and predictive representation, including for continuous or hybrid decisions.

We emphasize the scope of this statement. An abstract formulation is not evidence of empirical transfer. We demonstrate AGM through FORESIGHT on discrete road networks, and we do not claim that the present experiments establish effectiveness in other domains. Those extensions would require appropriate architectures, constraint definitions, and independent evaluation.

2.2 Relation to autoregressive generation and planning

AGM does not introduce a new probability chain rule or establish a strictly more expressive family of trajectory distributions. A sufficiently expressive autoregressive model can, in principle, represent the same conditional distribution. The difference is an explicit inductive bias and information-processing mechanism: an anticipated downstream representation participates in forming the current decision distribution, rather than downstream dependence being handled solely within a direct next-step predictor.

This design also shares a broad closed-loop intuition with receding-horizon planning, where future predictions are revised after actions are executed. The objectives, however, need not be the same. Conventional shortest-path optimization seeks a minimum-cost route under specified costs; FORESIGHT learns a distribution of observed route choices, including heterogeneous and non-optimal behaviors. The anticipated continuation can be viewed intuitively as learned guidance for a generative decision, but it is not an admissible A* cost-to-go heuristic and provides no optimality guarantee.

Therefore, the transferable insight we claim is the explicit separation and coupling of anticipation and current probabilistic decision-making within a repeated closed-loop generation process, not a universal advantage over autoregression or planning. We will sharpen this distinction and appropriately limit broader applicability claims in the revised manuscript.

3. Interpretation of Dijkstra and Ablation Results (CW3)

3.1 Why Dijkstra can be competitive

The reviewer correctly observes that Dijkstra can be competitive on some reported metrics, including against certain ablated variants. This is an important reference result. A shortest-path method has a strong structural prior and directly optimizes path cost; depending on the OD pairs and evaluation metric, a shortest path can resemble a reference trajectory closely even without learning from data.

However, a good cost-minimizing route is not equivalent to a learned distribution over route choices. For a fixed OD pair, deterministic Dijkstra does not itself represent the observed frequencies of alternative paths, heterogeneous preferences, or non-shortest behaviors. FORESIGHT targets these statistical properties while enforcing legal transitions. We therefore do not claim that FORESIGHT must outperform Dijkstra on every metric or every OD condition. Dijkstra is a meaningful optimization-based reference against which the added value of learned distribution modeling should be assessed.

3.2 What the ablation results do—and do not—establish

The relevant mechanisms in FORESIGHT serve different, coupled functions:

  • Future anticipation constructs an explicit predictive representation of downstream evolution available to the current transition decision.
  • The topology-constrained Transition Network transforms the current context and anticipation into probabilities over legal candidate movements.
  • Candidate-conditioned continuation learning couples the immediate transition probabilities to predictive errors along continuations associated with hypothetical legal movements.
  • Re-anticipation after execution refreshes the predicted downstream context using the newly realized history.

Consequently, removing or modifying one mechanism changes not only the information available to that mechanism but also the interactions through which the complete model is trained or used. For example, eliminating anticipation removes downstream predictive information from transition scoring; removing continuation-based learning eliminates a training signal that relates the immediate choice distribution to subsequent predictive behavior. Such variants should be understood as ablated configurations of a coupled architecture, not as separately optimized route-choice models.

The fact that some variants underperform Dijkstra is not evidence that Dijkstra and FORESIGHT solve identical modeling objectives, nor does it establish that every component individually improves every metric. It shows that retaining only part of the intended mechanism may not suffice to recover the complete model's empirical behavior. We will revise the experimental discussion to explain the specific functional role of each ablated mechanism and to avoid interpreting these results as proof of independent component-wise gains.

Q1. How do autoregressive methods, especially the basic Markov baseline, incorporate OD constraints?

Our evaluation distinguishes a common OD-conditioned testing protocol from the conditioning information used internally by each model. All methods are evaluated on the same fixed set of 5,000 test conditions, each specifying an origin, destination, and reference trajectory. Generation budgets and failure-accounting rules are aligned, and incomplete or unsuccessful generations are counted rather than replaced until a successful path is obtained.

The basic Markov baseline estimates local transition probabilities from training trajectories:

$ P(X_{t+1}\mid X_t), $

rather than explicitly modeling

$ P(X_{t+1}\mid X_t,g). $

It starts from the specified origin. Under the evaluation's terminal constraint, STOP is permitted only when the current state is the specified destination; arriving at that state does not by itself force an immediate stop. The destination does not enter the Markov baseline's intermediate next-state probability calculation. Thus, this baseline is a local-transition statistical reference, not a fully goal-conditioned autoregressive generator.

We agree that identical OD test conditions do not imply identical model-level conditioning capabilities. Comparisons against the basic Markov model should therefore not be interpreted as isolating the effect of anticipation from the effect of destination conditioning. We will make each baseline's inputs, transition restrictions, stopping behavior, and role in the comparison explicit in the revised evaluation description.

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

This paper studies the problem of trajectory generation, which is a very popular problem in spatiotemporal data mining domain. To address the problem, the authors develop FORESIGHT for destination-conditioned trajectory generation on road networks. The major difference between FORESIGHT applies an anticipate–select–update–re-anticipate loop rather than simulating location sequences directly. Experiments on real datasets show that FORESIGHT achieves state-of-the-art under some particular settings.

Critical strengths

CS1: It is a novel idea to enable a connection between future prediction and local transition selection. At an intersection, some legal movements may look plausible from the traversed history. Using an anticipated continuation to inform that choice, then recomputing it after moving, is a feasible approach to incorporating downstream context.

CS2: Technical details are well written. PyTorch style pseudocode for key algorithms are presented.

Critical weaknesses

CW1: The benefit of feeding anticipated futures into transition selection is not well supported by experiment results. The ablations remove future supervision, candidate compatibility, or inference-time re-ranking, but retain the connection from future representations to the Transition Network. In particular, setting the reranking weight to zero still leaves the base transition distribution conditioned on anticipated future representations. These comparisons show that the individual training and inference components help, but do not distinguish the proposed anticipation interface from a simpler model trained with auxiliary future prediction.

CW2: The authors claim that they advocate a fundamental shift: moving beyond location simulation toward behavior modeling on the road network, fundamentally shifting the focus from location simulation to route planning on road networks. However, destination-conditioned route generation and road-network-aware modeling are already developed in existing methods such as HOSER and Diff-RNTraj. Predicting future continuations to inform decisions, to some extent, also builds on existing ideas in sequence-model-based planning and receding-horizon control. The authors may want to elaborate their "shift" in a clearer way.

Questions for authors

Q1: Is it possible to compare the full model with a version that retains auxiliary future prediction but does not pass future representations into the Transition Network? Better to keep the state representation, graph constraints, and model capacity as comparable as possible.

Q2: How does the compatibility objective behave at a fork where alternative routes do not reconnect within the anticipation horizon?

7 comments

We sincerely thank the reviewer for the thoughtful and constructive comments. We particularly appreciate the recognition of the connection between future prediction and local transition selection, as well as the clarity of the technical presentation. We address the concerns regarding the anticipation mechanism, the positioning of our contribution, and the behavior of the compatibility objective below.

1. Response to CW1 and Q1: The Role of Anticipation in Transition Selection

We appreciate the reviewer's concern about whether the benefit of feeding anticipated futures into transition selection is sufficiently distinguished from that of auxiliary future prediction. We agree that this distinction is important for understanding the contribution of Anticipatory Generative Modeling (AGM).

FORESIGHT is not simply a conventional sequential generator augmented with an auxiliary future-prediction task. Rather, anticipation explicitly participates in the generative decision mechanism. The distinction lies in how the predicted future is incorporated into transition selection and how it is coupled with the learning objective.

In a conventional auxiliary-prediction formulation, the future decoder and transition network can be expressed as:

R^t=Fθ(ct,g),qt=Tϕ(ct,g;Ut).\widehat R_t=\mathcal F_\theta(c_t,g), \qquad q_t=\mathcal T_\phi(c_t,g;\mathcal U_t).

Although future prediction can provide additional supervision during training, the predicted future representation does not explicitly condition the transition distribution. Consequently, future prediction serves as an auxiliary learning signal, while the transition decision remains based on the observed history and destination.

In contrast, FORESIGHT explicitly conditions the legal transition distribution on the anticipated continuation:

$\widehat R_t=\mathcal F_\theta(c_t,g), \qquad q_t=\mathcal T_\phi(c_t,g,\widehat R_t;\mathcal U_t). $

Here, the anticipated continuation provides downstream predictive context for distinguishing among locally feasible movements. This is particularly important at branching points, where multiple legal transitions may be consistent with the observed history but lead to substantially different downstream trajectory evolutions.

Furthermore, the coupling between anticipation and transition selection extends beyond the forward computation. FORESIGHT introduces candidate-conditioned re-anticipation:

ct(s′)=ct⊕s′,R^t+1(s′)=Fθ(ct(s′),g),c_t^{(s')}=c_t\oplus s', \qquad \widehat R_{t+1}^{(s')}=\mathcal F_\theta(c_t^{(s')},g),

and incorporates the resulting continuation errors into the learning objective:

Lt=−mt,1log⁡qˉt(Yt,1)+∑s′∈Utmoveqt(s′)ℓtcomp(s′).\mathcal L_t =-m_{t,1}\log \bar q_t(Y_{t,1})+\sum_{s'\in\mathcal U_t^{\mathrm{move}}} q_t(s')\ell_t^{\mathrm{comp}}(s').

Importantly, candidate-conditioned continuation errors provide a learning signal for the transition distribution through their probability-weighted compatibility objective; they are not additional inputs used to form the forward transition scores. Thus, anticipation and transition selection are coupled at two complementary levels: anticipated representations condition the current transition decision, while candidate-conditioned continuation errors influence the learning of transition probabilities.

Regarding Q1, an auxiliary-prediction-only variant would indeed provide a comparison between two different ways of incorporating future prediction into trajectory generation. However, it changes the defining conditional structure of AGM rather than removing an optional component within the proposed formulation. In the auxiliary-prediction formulation, future prediction remains an additional learning objective without explicitly conditioning the transition distribution. AGM, instead, makes the predicted future an explicit decision context and further couples candidate-conditioned predictive compatibility with transition-probability learning. These operations constitute the proposed generative mechanism rather than independent auxiliary components.

Our existing ablations examine the roles of future supervision, candidate-conditioned compatibility, and inference-time consistency within FORESIGHT, while retaining the defining anticipation-to-transition dependency. We acknowledge that these experiments do not independently quantify the marginal performance contribution of this dependency against an auxiliary-prediction-only alternative. Rather, they evaluate the contributions of individual mechanisms within the integrated AGM formulation. We will clarify this distinction in the revised manuscript and avoid attributing to the existing ablations an isolated empirical comparison that they were not designed to provide.

2. Response to CW2: Distinction from Existing Trajectory Generation and Planning Methods

We appreciate the reviewer for pointing out the connections to existing destination-conditioned trajectory generation and planning approaches. However, we would like to clarify that the central contribution of FORESIGHT is not simply to incorporate road-network constraints, destination information, or future-oriented reasoning, but to establish a different generative mechanism for integrating explicitly predicted future information into graph-constrained transition decisions.

In particular, the distinction from existing methods lies in how downstream information is represented and incorporated into the generative process.

First, FORESIGHT differs from destination-oriented transition guidance in how future information is represented and used. For example, HOSER employs a Destination-Oriented Navigator that predicts the next reachable road segment by combining the current partial trajectory, destination information, and candidate-to-destination features. It then uses a search-based procedure to generate a trajectory between the given OD pair. We acknowledge that HOSER already incorporates destination guidance into local transition selection over reachable road candidates. However, FORESIGHT introduces a different mechanism: it explicitly predicts a multi-step, destination-conditioned future continuation R^t\widehat R_t, and uses its learned representation to inform the current transition distribution:

$\widehat R_t=\mathcal F_\theta(c_t,g),\qquad q_t=\mathcal T_\phi(c_t,g,\widehat R_t;\mathcal U_t).$

The distinction is that destination information indicates where the trajectory should eventually arrive, whereas the anticipated continuation provides a learned representation of how the trajectory may evolve downstream. This allows FORESIGHT to distinguish among locally feasible movements using predicted future behavioral patterns, rather than relying solely on the observed history and destination-oriented candidate features. In other words, destination guidance specifies the goal, while anticipation provides predictive context for deciding how to proceed toward that goal.

Second, global road-network-constrained trajectory generation differs from sequential anticipatory generation. Diff-RNTraj generates hybrid road-network trajectories through a continuous diffusion process and incorporates a spatial validity loss during denoising to encourage connectivity between adjacent road segments. FORESIGHT instead defines the next-transition distribution directly over the legal outgoing movements of the current road state:

supp⁡(qt)⊆Ut.\operatorname{supp}(q_t)\subseteq\mathcal U_t.

Its topological validity is therefore enforced through the support of the transition distribution, rather than through a validity objective applied to a globally generated trajectory representation. More importantly, FORESIGHT explicitly connects future prediction with each local transition decision, which is fundamentally different from generating a trajectory representation through a global diffusion process.

Third, although FORESIGHT shares the high-level predict–act–update pattern with receding-horizon planning, their underlying objectives and mechanisms differ. Conventional receding-horizon control typically predicts future system states to optimize a prescribed control objective and executes the first action of the resulting plan. FORESIGHT does not optimize a predefined route cost or commit to a fixed predicted route. Instead, it learns a conditional distribution over future continuations from observed trajectory data, uses the anticipated continuation to inform a probabilistic choice among legal movements, and recomputes the anticipation after each executed transition. Candidate-conditioned re-anticipation further couples these decisions with downstream prediction through the expected compatibility objective.

These distinctions are important because the contribution of FORESIGHT is not merely to apply destination guidance, road-network constraints, or receding-horizon execution to trajectory generation. Rather, it introduces an anticipation-conditioned generative decision process, in which an explicitly predicted future serves as an intermediate decision context for selecting the current topologically feasible movement. The anticipated future is provisional rather than a fixed plan, and the generation process repeatedly updates this context as the trajectory unfolds.

We therefore view the shift from location simulation toward behavior modeling as a methodological shift in the organization of the generative process. Rather than directly generating trajectory sequences or selecting local movements based only on the observed history and destination, FORESIGHT explicitly anticipates downstream trajectory evolution and incorporates this predictive information into graph-constrained generative decisions. This is the fundamental modeling distinction we intend to establish through AGM: future anticipation becomes an explicit and integral part of the generative decision process, rather than merely an auxiliary prediction task or an implicit consequence of sequential generation.

3. Response to Q2: Compatibility at Forks Without Reconnection

We thank the reviewer for raising this question. It is important to clarify that anticipation in FORESIGHT should not be interpreted as predicting a single future route that the generator is expected to follow, nor as requiring alternative candidate routes to eventually reconnect within the anticipation horizon.

Instead, anticipation provides a learned, destination-conditioned predictive context that summarizes possible downstream trajectory evolution. Its purpose is to guide the current transition decision, rather than to specify the exact future path. In this sense, it plays a role analogous to heuristic guidance in search: it provides information about downstream tendencies to inform the current decision, without committing the generator to a particular continuation. Unlike a conventional search heuristic, however, our anticipation is learned from trajectory data and represented through probabilistic future predictions and their latent representations. This analogy concerns the functional role of predictive guidance, rather than the mathematical form or optimality guarantees of conventional search heuristics.

Accordingly, at a fork where alternative routes do not reconnect within the anticipation horizon, FORESIGHT does not impose any requirement that their predicted continuations converge. Each candidate-conditioned re-anticipation,

R^t+1(s′)=Fθ(ct⊕s′,g),\widehat R_{t+1}^{(s')}=\mathcal F_\theta(c_t\oplus s',g), provides a candidate-specific predictive context under the corresponding candidate-extended history. Different candidates may therefore induce distinct future representations, reflecting their different downstream tendencies. This is expected behavior rather than a structural limitation of the framework.

The behavior of the compatibility objective at such forks requires a further distinction. For each candidate movement s′s', FORESIGHT evaluates its candidate-conditioned predictive distributions against the observed continuation of the current training trajectory. Thus, all candidate-conditioned predictions share the same observed suffix as their sample-conditioned training reference, even though unselected candidates may lead to different valid routes. The shared observed suffix serves as a common training reference; it does not assert that an unselected candidate would actually realize that same continuation.

Consequently, when an unselected candidate follows a different route that does not reconnect with the observed one within the anticipation horizon, its predicted continuation may have lower agreement with the observed suffix and incur a larger compatibility loss. Through the probability-weighted objective,

$\mathcal L_t^{\mathrm{comp}}=\sum_{s'\in\mathcal U_t^{\mathrm{move}}} q_t(s')\ell_t^{\mathrm{comp}}(s'),$

this difference provides a learning signal that favors candidates whose predicted continuations better explain the observed training sample.

More specifically, at a nonterminal fork, let zt(s′)z_t(s') denote the transition logit of candidate s′s', with qtq_t obtained through softmax normalization. Holding the candidate-conditioned compatibility losses fixed when taking this partial derivative, the direct contribution of the compatibility term to the transition-logit gradient is:

$ \frac{\partial \mathcal L_t^{\mathrm{comp}}} {\partial z_t(s')}=q_t(s') \left[ \ell_t^{\mathrm{comp}}(s')-\sum_{\bar s\in\mathcal U_t^{\mathrm{move}}} q_t(\bar s)\ell_t^{\mathrm{comp}}(\bar s) \right ]. $

This shows that candidates with compatibility losses above the current probability-weighted average receive a downward learning signal on their transition logits, whereas candidates with lower-than-average compatibility losses receive a relatively favorable signal. This expression isolates the direct logit-level contribution of the compatibility term; the complete objective also includes the realized next-event likelihood and the associated future-prediction supervision.

Reviewer_BTZzOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a modeling paradigm AGM, and its implementation Foresight, for road-network-constrained trajectory generation by anticipating potential branching points in advance. The method achieves empirical improvements.

Critical strengths
  1. The paper studies a problem with practical applications, focusing on trajectory generation under road network constraints.

  2. The paper proposes a new modeling approach that focuses on changes in branching points throughout the trajectory.

Critical weaknesses
  1. AGM lacks a rigorous definition. How does it formally differ from previous paradigms? If the difference is mainly in the motivation and interpretation, the authors should clarify this.

  2. Although the method achieves better results, the experimental section lacks sufficient analysis. Also, the ablation studies only cover some of the components, making it unclear where AGM's advantages over previous methods come from. Some interpretability analysis would be helpful.

Questions for authors
  1. In figure 1 most of the trajectories do not follow the roads, and some arrows have unclear meanings, making the figure somewhat misleading. The authors should consider improving this figure.

  2. AGM seems to be closely coupled with Foresight. Could the authors explain the relationship and differences between them?

  3. Figures 3 and 4 provide visualizations, but the comparisons are based on the simple Dijkstra baseline. The authors could consider including more advanced methods for comparison.

  4. Map-matching trajectories generated by other approaches, such as diffusion models, onto road networks seems to be a good solution within existing paradigms. In addition, many methods in autonomous driving and macroscopic trajectory generation have shown strong empirical results. Could the authors clarify what the limitations of these approaches are?

7 comments

Thank you for asking us to sharpen the definition of AGM and clarify what our experiments establish. We address these concerns through a more precise characterization of the modeling mechanism, quantitative ablation evidence, and expanded qualitative comparisons.

CW1 and Q2 — Definition of AGM and its relation to FORESIGHT.

AGM retains the standard sequential factorization of a trajectory distribution. Its distinction lies in how each transition distribution is constructed and learned.

Given the realized prefix ctc_t and destination gg, AGM explicitly predicts a provisional future continuation R^t=Fθ(ct,g)\hat R_t=\mathcal F_\theta(c_t,g) and incorporates its predictive representation into the current transition distribution qt=Tϕ(ct,g,R^t;Ut)q_t=\mathcal T_\phi(c_t,g,\hat R_t;\mathcal U_t). After executing a transition, the model updates the realized prefix and recomputes the anticipation.

A conventional direct-transition model can access the same history, destination, and feasible candidate set. However, AGM explicitly constructs predictive information about downstream evolution before making the current transition decision. This anticipation is neither an observed future nor a committed route, but a learned predictive representation that informs the current decision.

The mechanism shares the closed-loop structure of receding-horizon planning, but our objective is to learn a generative distribution over legal route choices from observed trajectories, rather than to optimize a deterministic planning cost.

FORESIGHT realizes this principle through a finite-horizon Future Anticipation Decoder, a topology-constrained Transition Network, and candidate-conditioned continuation learning. Thus, AGM specifies the modeling principle, while FORESIGHT provides its concrete implementation for road-network-constrained trajectory generation.

CW2 — What the ablations establish.

FORESIGHT connects anticipation and transition selection through complementary training- and inference-time mechanisms.

During training, candidate-conditioned forecasts are computed after the base transition distribution qtq_t has been formed. Each hypothetical movement is evaluated against the same observed suffix through a masked continuation loss, with candidate-specific costs weighted by qtq_t. This provides a shared-reference compatibility objective, not ground-truth supervision for hypothetical counterfactual trajectories. During inference, the optional reranking mechanism instead compares overlapping model forecasts to construct the final selection distribution πt\pi_t (Eqs. 10–11).

Table 2 and Appendix E.2 provide quantitative evidence for these mechanisms relative to each city's complete-history FORESIGHT-F.

On Porto, Distance JSD increases from 1.53 to 2.55 when candidate-compatibility learning is removed, to 2.25 when future-window supervision is removed, and to 2.05 when inference-time reranking is disabled (all values in units of 10−210^{-2}; lower is better).

The first two variants are independently retrained, whereas the reranking ablation uses the same checkpoint with w=0w=0. The reported point estimates exhibit the same ordering in San Francisco and Beijing.

These comparisons demonstrate the contributions of the two training objectives and inference-time reranking within the FORESIGHT architecture. They do not independently establish the causal contribution of explicit anticipation relative to every possible direct-transition architecture; such attribution would require a matched direct-transition control. We therefore distinguish the demonstrated component-level improvements from broader claims about the AGM modeling principle.

Q1 and Q3 — Improved qualitative visualizations.

We have revised Figure 1 to clarify the distinction between prediction and execution. Purple dashed arrows represent provisional future anticipations, whereas solid blue arrows represent executed graph-legal transitions. The red curve in panel (a) is a conceptual illustration of unconstrained generation, rather than an output attributed to a particular baseline. These visual conventions are clarified in the figure caption.

We have also expanded Figures 3 and 4 by including HOSER and Diff-RNTraj, alongside Dijkstra as a deterministic shortest-path reference.

  • Figure 3 compares aggregate spatial trajectory patterns in Porto, illustrating the distribution of generated movements over the road network.
  • Figure 4 compares generated routes under two fixed test OD conditions, illustrating route-level differences among the methods.

These additions extend the original shortest-path comparison to representative learning-based navigational and diffusion-based trajectory generators. The two figures provide complementary distribution-level and route-level perspectives, without treating individual route examples as sufficient evidence of distributional fidelity.

Q4 — Map matching and related generative approaches.

We agree that diffusion followed by map matching constitutes a viable approach to generating road-network trajectories. Our comparison does not assume that diffusion models are inherently incapable of producing topologically valid routes.

Table 1 already includes an OD-adapted DiffTraj baseline, whose generated GPS trajectories are converted to the common road representation through a fixed map-matching procedure (Appendix D.1). On Porto, DiffTraj achieves a Distance JSD of 2.83 and EDR of 0.436, compared with 1.53 and 0.312 for FORESIGHT-F under the reported evaluation protocol.

Map matching can enforce network connectivity, but its resolution of ambiguous paths may also affect the resulting route distribution. In contrast, FORESIGHT assigns probabilities directly over legal graph transitions, using anticipated downstream information to inform selection among feasible alternatives. The expanded qualitative comparisons additionally include the road-aware diffusion model Diff-RNTraj.

Our evaluation concerns complete, OD-conditioned routes over directed road networks. Autonomous-driving planners may share the anticipation and replanning principle but typically address additional vehicle dynamics, collision avoidance, and interactions. Macroscopic mobility generation may emphasize aggregate flow statistics, whereas our evaluation additionally examines individual route fidelity under matched OD conditions.

These distinctions define the scope of our current implementation rather than excluding the potential applicability of anticipation-informed generation to continuous or hybrid decision spaces.

I appreciate authors' response. However, it does not address my major concerns. Foresight restricts the available actions to achieve on-road trajectories, whose comparison with the unconstrained methods is incremental and insufficient to serve as an explanation of motivation. The rationality of this paradigm also lacks sufficient justification based on theoretical or empirical evidence. After considering the opinions of other reviewers and authors' responses, I think the work exaggerates its contributions. Therefore, I lower my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Consolidated Response and Summary of Clarifications

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate their recognition of the practical relevance of the problem, the methodological design of FORESIGHT, and its empirical performance. The discussion has helped us clarify the definition of Anticipatory Generative Modeling (AGM), its relationship to existing approaches, and the evidence provided by our experiments.

1. Core contribution: anticipation as part of generative decision-making.

AGM does not introduce a new probability factorization, nor does it claim that destination-conditioned route choice is a new problem. Its contribution is a different organization of the generative decision process: explicitly predicting downstream evolution before constructing the current transition distribution, then updating that prediction as generation proceeds.

FORESIGHT implements this principle by conditioning graph-constrained transition selection on an anticipated future representation, coupling transition-probability learning with candidate-conditioned continuation prediction, and re-anticipating after each executed movement. Thus, FORESIGHT is not simply a sequential generator with an auxiliary future-prediction task: anticipated future information directly informs the current decision. The anticipation is provisional guidance, not a predetermined route or a ground-truth counterfactual.

2. Relationship to prior work and scope.

We acknowledge that classical route-choice models, destination-conditioned generators, and planning methods already use downstream or destination information. We therefore do not claim to introduce future-aware route selection for the first time. Rather, FORESIGHT integrates explicit multi-step anticipation, probabilistic selection over legal graph transitions, candidate-conditioned continuation learning, and closed-loop re-anticipation into one generative mechanism.

Unlike deterministic shortest-path optimization, our objective is to learn a distribution over observed route choices, including plausible non-shortest alternatives. Unlike global trajectory generation, FORESIGHT constructs trajectories through successive topologically feasible decisions informed by provisional future predictions. We also distinguish AGM as a modeling principle from FORESIGHT as its road-network-specific realization. Although the principle may extend to other decision spaces, our experiments do not establish empirical generalization beyond road-network trajectories.

3. Experimental evidence and additional analysis.

Our experiments evaluate the complete FORESIGHT framework, while the ablations examine future-window supervision, candidate-conditioned compatibility learning, and inference-time future-consistency reranking within the integrated architecture. These ablations assess complementary mechanisms; they do not independently establish the causal contribution of every design choice against all possible alternative architectures.

We have also revised Figure 1 to distinguish provisional anticipation from executed transitions, and expanded Figures 3 and 4 with representative learning-based baselines, including HOSER and Diff-RNTraj. The additional sensitivity analysis shows relatively stable, metric-dependent behavior across the tested inference-time consistency weights. Together, these analyses provide complementary quantitative and qualitative evidence without overstating what individual comparisons establish.

4. Clarifications to evaluation and interpretation.

We further clarify how candidate-conditioned compatibility learning operates at branching points, including when alternative routes do not reconnect within the anticipation horizon. We distinguish shared OD test conditions from the actual conditioning capabilities of individual baselines, explain the role of map matching in diffusion-based comparisons, and address data privacy considerations and limitations of the current evaluation.

In summary, the central contribution is an integrated generative mechanism in which explicit future anticipation actively informs current probabilistic decisions, rather than serving only as an auxiliary prediction objective. FORESIGHT provides a concrete realization of this principle for destination-conditioned, road-network-constrained trajectory generation.

We believe these clarifications, analyses, and visualizations more precisely establish the methodological contribution, strengthen the interpretation of the empirical results, and delineate the scope of our claims. We sincerely appreciate the reviewers' feedback and the opportunity to improve our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.