冻结扩散模型的可识别世界模型：Contrastive Diffusion Alignment 方法

View PDF HTML (experimental)

Abstract:Diffusion-based world models can generate and predict trajectories in high-dimensional dynamical systems, but predictive accuracy does not imply that their latent coordinates recover the underlying state variables or causal interactions. We ask whether a frozen pretrained diffusion model can be equipped with identifiable coordinates without retraining its generative backbone. We show that auxiliary-variable nonlinear ICA guarantees can be transferred to Contrastive Diffusion Alignment (ConDA), which learns only a lightweight alignment map on top of frozen diffusion latents. Under standard TCL/GCL assumptions, the aligned representation identifies latent dynamical states up to permutation and componentwise invertible transformations, preserves the latent dynamic structural causal model, and reduces lagged graph recovery to transition-Jacobian sparsity. We evaluate TCL-, GCL-, and CEBRA-based ConDA against TDRL, CaRiNG, IDOL, temporal SuaVE, and iVAE across physical and robotic video systems. TCL and GCL achieve near-perfect blockwise state recovery and competitive lagged graph recovery, including exact recovery in a simulated falling-body system. In a simulated bipedal robot, learned dynamics recover the sign and temporal structure of responses to held-out control perturbations. These results show that a frozen generative diffusion model can be equipped with coordinates that are identifiable, structurally interpretable, and useful for analyzing intervention-relevant dynamics.

Subjects:	Machine Learning (cs.LG)
Cite as:	arXiv:2610.07028 [cs.LG]
	(or arXiv:2610.07028v1 [cs.LG] for this version)
	https://doi.org/10.48550/arXiv.2610.07028 arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Ruchi Sandilya [view email]
[v1] Sun, 4 Oct 2026 18:15:13 UTC (865 KB)