acceptodds
Withdrawn from ICLR 2027

Bridging Heterogeneous Spaces via Latent Asymmetric Bidirectional Diffusion Bridge for Human Mobility Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion bridges have demonstrated remarkable success in paired translation tasks by constructing a continuous stochastic process between two coupled distributions. However, existing frameworks intrinsically rely on the assumption of homogeneous boundary domains (i.e., an assumption that requires identical topologies and dimensionalities), which renders these frameworks mathematically ill‑posed when mapping across heterogeneous state spaces. In this paper, we propose the Latent Asymmetric Bidirectional Diffusion Bridge (LAB-Bridge), a novel theoretical framework designed to construct a reversible Markov process across topologically mismatched spaces. By projecting disparate modalities into a harmonized latent Riemannian manifold via diffeomorphic mappings, we formally extend the Chapman-Kolmogorov Equations to heterogeneous boundaries. We provide a rigorous bi-Lipschitz bound guaranteeing the preservation of intrinsic topologies during this projection. Furthermore, we mathematically derive the continuous-time bidirectional Stochastic Differential Equations (SDEs) and prove the analytical equivalence of their drift fields. This equivalence enables the derivation of a highly efficient, unified asymmetric score estimator. We establish a theoretical bound demonstrating that our single-network masking paradigm optimally minimizes the Kullback-Leibler divergence of path measures in both directions simultaneously. We empirically validate LAB-Bridge on complex cross-manifold translation tasks. Specifically, we abstract human mobility dynamics as a heterogeneous state-to-state mapping challenge, demonstrating that our framework effectively translates between discrete historical preference sequences and continuous point-of-interest spatial distributions, achieving state-of-the-art bidirectional generative fidelity.

Primary area:
learning on time series and dynamical systems
Keywords:
Human mobility modeling, Diffusion bridges, Heterogeneous state‑space translation, Latent Riemannian manifold, Point‑of‑interest recommendation

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_BMP1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the Latent Asymmetric Bidirectional Diffusion Bridge (LAB-Bridge) for translation across heterogeneous data domains. Modality-specific encoders map both domains into a common latent space, where an entropy-regularized optimal transport objective aligns their distributions. A shared score network, conditioned on either endpoint via Bernoulli masking, supports bidirectional generation, with modality-specific decoders mapping generated latent states back to the respective domains.

The paper presents theoretical claims regarding topology preservation, bidirectional transition kernels, stochastic differential equations (SDEs), and the relationship between the training objective and path-space KL divergence. The framework is applied to human mobility modeling, translating discrete historical preference sequences to continuous geographic coordinates and vice versa. On five Foursquare city datasets, the authors report improvements over recommendation and diffusion baselines in spatial prediction and discrete ranking metrics. Additional evaluations include component ablations, latent geometric distortion measurements, and bidirectional cycle-consistency comparisons.

Critical strengths
  1. The paper addresses a meaningful problem: bidirectional translation between discrete mobility histories and continuous spatial information.
  2. Experiments across five city datasets report consistent improvements and include component ablations, providing encouraging empirical evidence, subject to the evaluation concerns below.
Critical weaknesses

The transition kernels and SDEs are not established by the stated probabilistic arguments.

In Appendix A.2, the step from Eq. (25) to Eq. (26) assumes that the noise terms vanish after conditioning on zt,z0z_t,z_0. However, ztz_t depends on ϵt\epsilon_t, and E[ϵt∣zt,z0]\mathbb{E}[\epsilon_t\mid z_t,z_0] is generally nonzero. For example, fix z0=0z_0=0, let zT,ϵt,ϵsz_T,\epsilon_t,\epsilon_s be independent standard Gaussian variables, and use Eqs. (22)–(23). Gaussian conditioning gives

E[zs∣zt,z0=0]=βsβtβt2+σt2zt,\mathbb{E}[z_s\mid z_t,z_0=0] = \frac{\beta_s\beta_t}{\beta_t^2+\sigma_t^2}z_t,

rather than (βs/βt)zt(\beta_s/\beta_t)z_t as in Eq. (26). The analogous conditional-expectation step in Eq. (29) has the same issue.

There is also a mismatch between the theorem statement, which conditions on both endpoints, and the proof, which switches to conditioning on only one endpoint. Algebraically eliminating an endpoint is not equivalent to marginalizing over its conditional distribution. Under the independent-noise construction used in the proof, conditioning on both endpoints makes zsz_s and ztz_t independent, yielding a forward conditional mean of αsz0+βszT\alpha_s z_0+\beta_s z_T, inconsistent with Eq. (20). More generally, the one-time marginals alone do not uniquely determine the temporal coupling or transition kernels.

Appendix A.3 introduces a further issue: Eq. (36) treats q(zt∣z0)q(z_t\mid z_0) as Gaussian with covariance σt2I\sigma_t^2I, although the specified Gaussian distribution conditions on both endpoints. Marginalizing over zTz_T generally yields a Gaussian mixture, not necessarily a single Gaussian, with conditional covariance

Cov⁡(zt∣z0)=σt2I+βt2Cov⁡(zT∣z0).\operatorname{Cov}(z_t\mid z_0) = \sigma_t^2I+\beta_t^2\operatorname{Cov}(z_T\mid z_0).

Thus, the score expression in Eq. (36) does not follow from the stated assumptions. Separately, the coefficient compatibility required to equate Eqs. (34) and (37), and the score identity needed to obtain Eq. (40) from Eq. (39), are not established.

These are central rather than peripheral issues: Theorem 1 is presented as the theoretical cornerstone of the method, Theorem 2 derives the sampling dynamics from its transition kernels, and the inference algorithms use those dynamics. Consequently, the current manuscript does not establish its main claimed theoretical contribution. A satisfactory resolution would require a corrected probabilistic construction and a demonstration that the resulting dynamics agree with the implemented sampler. The reported empirical improvements do not, by themselves, resolve these theoretical concerns.

Questions for authors
  1. What temporal coupling or reference process defines the transition kernels in Theorem 1? Please clarify whether Eqs. (20)–(21) condition on both endpoints or only one, and explain how Eqs. (26) and (29) account for the nonzero conditional expectations of the noise terms identified above.

  2. Can you provide a corrected derivation connecting the bridge marginals to the SDEs in Theorem 2? In particular, please justify the single-endpoint score in Eq. (36), establish the compatibility conditions required for Eqs. (34)–(37), and derive Eq. (40) from the time-reversal identity in Eq. (39). Do the resulting dynamics and schedules agree with those used in the implemented sampler?

  3. What exactly are the inputs, outputs, prediction horizons, and data splits for each evaluation direction? Please reconcile next-destination prediction with the trajectory-based ADE/FDE definitions, and historical-sequence recovery with the single-POI ranking metrics. How were the baseline methods adapted to these tasks?

Reviewer_ZL33OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces LAB-Bridge, a framework for diffusion bridges between topologically heterogeneous state spaces, where existing bridge methods (Schrodinger bridges, denoising diffusion bridges, rectified flows) are ill-posed because they assume identical source and target domains. The method projects both modalities into a shared latent Riemannian manifold via bi-Lipschitz diffeomorphic encoders aligned with Sinkhorn-regularized optimal transport, and constructs the bridge in this latent space. The authors prove a distortion bound, tying topological fidelity to the latent transport cost, derive exact bidirectional continuous-time SDEs with analytically equivalent forward and backward drifts, and show that a single masked score network upper-bounds the path-measure KL divergence in both temporal directions. On five Foursquare city datasets, LAB-Bridge achieves state-of-the-art results in both directions (up to 56.6% lower ADE forward, +14.29 points HR@10 backward on Sao Paulo), while zero-padded homogeneous bridges degrade.

Critical strengths
  1. This is a genuinely well-posed problem. The homogeneous domain assumption in diffusion bridges is a fundamental mathematical issue, which also arises when estimating common information-theoretic quantities via bridge matching. The paper formalizes the resulting ill-posedness and replaces zero padding with a principled latent harmonization construction, which is a notable contribution.
  2. The evaluation protocol is rather thorough within the geospacial domain. There are five cities, both directions, diverse strong baselines including forced-homogeneous variants. The negative result for zero-padded bridges (underperforming plain diffusion) supports the core claim.
Critical weaknesses
  1. The evaluation scope is narrow. There is one application domain, and a greedy decoding is used for the backward task. Claims about other heterogeneous pairs (e.g., graph to molecule) are only stated in the broader impact section, and are not tested.
Questions for authors
  • Can you evaluate your method on other domains where the dimensionality mismatch is also natural? E.g., for bridge matching-based mutual information estimation [1] between XX and YY of different dimensionalities?
  • Have you tried noise padding as a baseline alternative to zero padding?
  • What is the "maximum span" factor in the final step of the Proposition 1 proof, and can it be stated explicitly?
  • Can the backward score sign flip in Appendix A.3 be derived explicitly, rather than invoked via topological symmetry?
  • Proposition 2 gives an upper bound on the path-measure KL divergence. Does the "optimal" claim in the abstract hold?

[1] Kholkin et al. "InfoBridge: Mutual Information estimation via Bridge Matching", ICLR 2026.

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

The authors propose a method of bridging between heterogeneous spaces, nmotivated and applied to human mobility modelling, by projecting both spaces into a shared latent Riemannian manifold using diffeomorphic encoders.

Critical strengths
  1. The paper proposes an interesting approach towards bridging heterogeneous domains, and assesses the limitations in current flow-based and diffusion bridge methods.
  2. The proposed method outperforms state of the art methods (with some doubts. See weaknesses and questions)
Critical weaknesses

1.The authors oversell the theoretical novelty of bridging heterogeneous spaces. The idea is already used in latent diffusion model. The method can be explained as 2 diffeomorphic autoencoders into a shared latent space. 2. I am unfamiliar with the domain specific problem, but only single runs are reported, and the difference against other methods do not seem impressive, althoiugh this may be me not understaning the specific metric. 3. The compared against methods use zero-padding, and I doubt that makes up for a fair comparison.

Questions for authors
  1. Related to results, is there a dairer way to compare that is not zero-padding approaches?
  2. Do the results hold statistically. I mean, during multiple runs, is there statistically enough evidence that the method outperforms state of the art, or were there "lucky" runs?
  3. What differenciates this method from other approaches exactly? Is the primary contribution the Sinkhorn alignment in latent space, and if so, how does this differ from existing optimal transport regularized latent spaces?

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

Related papers

Discussion (0)

Sign in to comment.