Equivariant Neural Belief Propagation

A new factor-graph framework called Equivariant Neural Belief Propagation (ENBP) performs probabilistic inference over spatially embedded variables with high accuracy while respecting SE(3) symmetry, its developers report. The model synthesizes rank-2 precision tensors that existing equivariant networks cannot produce, enabling anisotropic uncertainty estimates [1]. The framework, detailed in a paper submitted to arXiv on 4 June 2026, represents messages as equivariant Gaussian mixture models whose sufficient statistics transform exactly under SE(3) [1]. Rank-2 precision matrices are built through equivariant outer products and ingested via differentiable spectral decomposition. A greedy KL-based mixture reduction keeps the representation tractable and provably commutes with SE(3) [1]. On the GEOM-QM9 and GEOM-Drugs molecular datasets, ENBP achieved 98.9% conformational coverage at 0.090 ångström error with sub-second latency [1]. The authors report it is over 100 times faster than diffusion baselines while delivering higher accuracy [1]. In multi-body robotic inference tests, standard loopy belief propagation diverged when 15 or more agents were present, whereas ENBP converged with near-zero collision rates [1]. The framework recorded equivariance error of approximately 10⁻⁷, compared with 10⁻¹ for augmented baselines [1]. Belief propagation operates on factor graphs, data structures that represent probability distributions over variables. Prior work has characterized how several indices in a factor graph can be permuted without changing the underlying distribution, and proposed Factor-Equivariant Neural Belief Propagation (FE-NBP) and Factor-Equivariant Graph Neural Networks (FE-GNN) to exploit this inductive bias [3]. Those models achieved state-of-the-art performance on small and large datasets respectively for both marginal and maximum a posteriori inference [3]. Equivariance — the property that a model’s output transforms predictably when its input is rotated or translated — has been a central design principle in deep learning. Convolutional neural networks, for instance, provide translation-equivariant responses through filters that slide across input features, though most are not strictly invariant to translation due to downsampling operations [4]. ENBP extends equivariant design to the precision tensors required for anisotropic uncertainty, a capability the authors note is absent from existing equivariant networks that produce only scalars and vectors [1].