Spatially Mapped Belief Masses

• Spatially mapped belief masses are probability distributions anchored to explicit spatial representations, integrating geometry and topology into the inference process.
• They are applied in robotics, molecular modeling, and 3D scene understanding, using methods like equivariant Gaussian mixtures and voxel-based mapping to update beliefs with spatial evidence.
• These approaches preserve multi-modal uncertainty and anisotropic characteristics, overcoming traditional latent state limitations by coupling belief updates with physical spatial structure.

Spatially mapped belief masses are probability distributions whose support, sufficient statistics, or marginals are tied explicitly to a spatial representation rather than stored only in an abstract latent state. In current arXiv work, this idea appears in several technically distinct forms: equivariant Gaussian-mixture messages over variables in $\mathbb{R}^3$, explicit 3D fields of Gaussian scene primitives, factorized beliefs over metric-semantic voxel maps, and lifted message multiplicities that stand in for many equivalent ground messages in a structured graph (Cheng et al., 4 Jun 2026, Yin et al., 12 May 2026, Marques et al., 28 Feb 2025, Smith et al., 2016). Taken together, these formulations treat belief as something that is mapped onto geometry, topology, or graph structure, so that inference updates a spatially grounded probabilistic object rather than a point estimate.

1. Formal meaning and representational scope

The common thread is the replacement of monolithic hidden-state inference with belief representations whose elements are individually anchored to spatial locations, geometric variables, or structured regions. In Equivariant Neural Belief Propagation (ENBP), the latent variables are spatial coordinates $\mathbf{x}_i \in \mathbb{R}^3$ with joint distribution

$$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$

and the message itself is a spatial probability distribution over $\mathbf{x}_i$. In 3D-Belief, the relevant belief is the POMDP posterior

$b(s^t) = P(s^t \mid o^{1:t}),$

but it is instantiated not as a monolithic hidden state but as a queryable 3D representation $z^t=\phi(s^t)$. In map-space belief prediction for manipulation-enhanced mapping, the belief is summarized by a metric-semantic voxel map and factorized across cells as

$$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$

A broader structural analogue appears in lifted region-based belief propagation, where a single symbolic belief or message can represent many equivalent ground instances through multiplicity counts (Cheng et al., 4 Jun 2026, Yin et al., 12 May 2026, Marques et al., 28 Feb 2025, Smith et al., 2016).

Work	Spatial support	Belief mass form
ENBP	$\mathbb{R}^3$ variables on a factor graph	Equivariant Gaussian mixture messages
3D-Belief	3D Gaussian Splatting scene representation	Gaussian primitives over observed and imagined scene regions
Map-space belief prediction	Metric-semantic voxel grid and BEV map	Per-cell occupancy and semantic marginals
LGBP	Lifted region graph	Symbolic messages with multiplicity factors

This suggests a unifying interpretation: a spatially mapped belief mass is not merely a posterior over abstract states, but a posterior whose internal coordinates are themselves meaningful for geometry, planning, or structured inference.

2. Continuous-space belief masses in three dimensions

ENBP gives the most explicit continuous-space formulation. Each factor-to-variable message is an Equivariant Gaussian Mixture Model (EGMM),

$$m_{a \to i}(\mathbf{x}_i) = \sum_{k=1}^{K} w_k \, \mathcal{N}\!\left(\mathbf{x}_i \,\middle|\, \boldsymbol{\mu}_k,\, \boldsymbol{\Lambda}_k^{-1}\right),$$

where $w_k$ are invariant mixing weights, $\mathbf{x}_i \in \mathbb{R}^3$0 are $\mathbf{x}_i \in \mathbb{R}^3$1-equivariant means, and $\mathbf{x}_i \in \mathbb{R}^3$2 are $\mathbf{x}_i \in \mathbb{R}^3$3-equivariant precision tensors. Under rotation $\mathbf{x}_i \in \mathbb{R}^3$4, the means and precisions transform as

$\mathbf{x}_i \in \mathbb{R}^3$5

and the density satisfies $\mathbf{x}_i \in \mathbb{R}^3$6 for the correspondingly transformed parameters. The paper’s central point is that beliefs are literally anchored in geometry and move with the world rather than with the coordinate frame (Cheng et al., 4 Jun 2026).

3D-Belief adopts a different continuous 3D representation. Its scene belief state is

$\mathbf{x}_i \in \mathbb{R}^3$7

where each Gaussian primitive stores a mean $\mathbf{x}_i \in \mathbb{R}^3$8, covariance $\mathbf{x}_i \in \mathbb{R}^3$9, opacity $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$0, spherical-harmonic appearance coefficients $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$1, and semantic embedding $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$2. The scene is split into observed and imagined components,

$$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$3

Here the mapped belief mass is a 3D field of Gaussian primitives distributed over space, with the imagined component encoding hypothesized unseen content (Yin et al., 12 May 2026).

The technical distinction between these two formulations is important. ENBP uses Gaussian mixtures as messages in a factor-graph inference algorithm; 3D-Belief uses Gaussian primitives as the scene representation over which a generative world model performs sequential inference. In both cases, however, spatial uncertainty is carried by explicit geometric objects rather than by latent feature vectors alone.

3. Symmetry, anisotropy, and the geometry of uncertainty

A defining issue in ENBP is that probabilistic inference over spatially embedded variables must respect rigid-motion symmetry. The paper states that physical systems are $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$4-invariant,

$$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$5

and argues that standard $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$6-equivariant GNNs are insufficient because they produce scalars and vectors, not the rank-2 precision tensors required for directional uncertainty. This matters when uncertainty is anisotropic, as in molecular bonds, torsions, and collision constraints (Cheng et al., 4 Jun 2026).

To construct these tensors, ENBP synthesizes precisions from equivariant vectors by

$$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$7

with $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$8 via softplus, $$p(\mathbf{x}) = \frac{1}{Z} \prod_{a=1}^{M} \psi_a(\mathbf{x}_{\mathcal{N}(a)}),$$9 equivariant basis vectors, and $\mathbf{x}_i$0 an isotropic regularizer ensuring positive-definiteness. The paper proves positive-definiteness, $\mathbf{x}_i$1-equivariance of the construction, and that $\mathbf{x}_i$2 can represent full-rank tensors. It also uses natural parameters

$\mathbf{x}_i$3

because Gaussian products are additive in natural-parameter space (Cheng et al., 4 Jun 2026).

3D-Belief does not frame its geometry through group-equivariant message passing, but it likewise treats uncertainty as inherently spatial and structured. Its representation preserves geometry, supports semantic querying, and distinguishes observed memory from imagined hypotheses. A plausible implication is that the two papers address different but compatible aspects of mapped belief: ENBP formalizes symmetry-respecting local belief transport, whereas 3D-Belief formalizes explicit scene-level belief maintenance under partial observability (Yin et al., 12 May 2026).

4. Belief updating under observations, actions, and message passing

The update mechanisms differ substantially across the cited systems, but all are designed to revise a spatially explicit belief object. In ENBP, variable-to-factor products use Gaussian natural-parameter addition: $\mathbf{x}_i$4 The formal factor-to-variable update is

$\mathbf{x}_i$5

and ENBP approximates it with an $\mathbf{x}_i$6-equivariant network $\mathbf{x}_i$7 that ingests spectral decomposed incoming precision tensors and relative coordinates $\mathbf{x}_i$8, then outputs $\mathbf{x}_i$9 for the outgoing EGMM. Final marginals are obtained by

$b(s^t) = P(s^t \mid o^{1:t}),$0

This is belief propagation in the literal sense: the mapped mass is transported and recombined across a factor graph (Cheng et al., 4 Jun 2026).

In 3D-Belief, the update is sequential and generative. The POMDP belief evolves as

$b(s^t) = P(s^t \mid o^{1:t}),$1

while the structured 3D belief state is updated autoregressively as

$b(s^t) = P(s^t \mid o^{1:t}),$2

The observed component expands with new evidence, producing $b(s^t) = P(s^t \mid o^{1:t}),$3, and the imagined component is replaced by a new imagination $b(s^t) = P(s^t \mid o^{1:t}),$4. The paper stresses that prior hallucinated content is not preserved as truth; it is treated as a revisable hypothesis. The update can be performed online with constant per-step cost (Yin et al., 12 May 2026).

Map-space belief prediction introduces Calibrated Neural-Accelerated Belief Updates (CNABUs) for two update regimes: $b(s^t) = P(s^t \mid o^{1:t}),$5 for viewpoint updates and $b(s^t) = P(s^t \mid o^{1:t}),$6 for manipulation updates. The observation network outputs evidential parameters $b(s^t) = P(s^t \mid o^{1:t}),$7 for semantic Dirichlet distributions and $b(s^t) = P(s^t \mid o^{1:t}),$8 for voxelwise Beta occupancy distributions, with beliefs given by

$b(s^t) = P(s^t \mid o^{1:t}),$9

The manipulation network also predicts $z^t=\phi(s^t)$0, a Beta distribution over whether each voxel changed due to the action. The push is encoded by 2D binary masks for start point, end point, and swept volume, so that the update predicts how the map itself will change before the next observation (Marques et al., 28 Feb 2025).

5. Multimodality, calibration, and tractable aggregation

A recurrent theme is that mapped belief masses must represent ambiguity without collapsing it into physically implausible averages. ENBP states directly that single-Gaussian messages are too restrictive because they collapse multi-modal energy landscapes to a mean that may have little physical probability. To keep multi-component messages tractable, the method greedily merges components using a KL-based criterion following Runnalls: it chooses the pair with the smallest closed-form KL upper bound and replaces them by a moment-matched merged component. The paper emphasizes that this reduction is equivariance-preserving because the KL criterion is built from invariants such as trace, quadratic form, and log-det, and because moment matching commutes with rotation (Cheng et al., 4 Jun 2026).

3D-Belief handles ambiguity through generative sampling over entire 3D scenes. Its scene-level diffusion model operates over 3D Gaussian primitives rather than over video frames, and uncertainty in unseen space is expressed by sampling alternative 3D Gaussian scenes from a conditional diffusion posterior. Multi-hypothesis sampling is explicit: in planning, the model uses 3 hypotheses per decision step. This makes the belief state actionable without requiring commitment to a single completion of partially observed space (Yin et al., 12 May 2026).

Map-space belief prediction addresses a different failure mode: overconfidence in unknown or manipulated regions. Its evidential losses regularize predictions toward a non-informative prior unless supported by data: $z^t=\phi(s^t)$1 The paper presents this as the key calibration mechanism for meaningful information-gain analysis. For manipulation updates it also includes a consistency loss to keep evidence stable where the map should not change (Marques et al., 28 Feb 2025).

LGBP addresses tractability at the graph-structural level rather than the geometric level. Instead of sending one message per ground edge, it computes how many copies of each message exist in the simulated ground graph and reuses one symbolic message with multiplicities such as $z^t=\phi(s^t)$2, $z^t=\phi(s^t)$3, and $z^t=\phi(s^t)$4. In this formulation, one symbolic message stands for many equivalent ground messages, and the exponents encode repeated belief-mass contributions (Smith et al., 2016).

6. Empirical behavior, uses, and limitations

The practical value of spatially mapped belief masses is most explicit in ENBP’s molecular and robotic experiments. On GEOM-QM9 and GEOM-Drugs, at generation budget 200, ENBP reports QM9 Coverage $z^t=\phi(s^t)$5, QM9 AMR $z^t=\phi(s^t)$6 $z^t=\phi(s^t)$7, Drugs Coverage $z^t=\phi(s^t)$8, Drugs AMR $z^t=\phi(s^t)$9 $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$0, and Time $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$1 s. In multi-body robotic inference, vanilla loopy BP diverges for $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$2, whereas ENBP converges even at $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$3; on the 20-agent 3D task, the reported equivariance errors are $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$4 for full ENBP, $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$5 for a Direct GNN (aug.), and $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$6 for Vanilla Loopy BP. At $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$7, 3D, ENBP with $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$8 gives NLL $$b(s) = \prod_{m_{i,j,k} \in M} P(m^{i,j,k} = s^{i,j,k}).$$9 and collision rate $\mathbb{R}^3$0, while ENBP with $\mathbb{R}^3$1 has a much worse collision rate of $\mathbb{R}^3$2. The reported interpretation is that multi-modal, anisotropic, symmetry-respecting belief masses are necessary for stable and physically plausible inference (Cheng et al., 4 Jun 2026).

3D-Belief evaluates mapped belief quality through three families of tasks: 2D scene memory and imagination, 3D imagination on the proposed 3D-CORE benchmark, and open-vocabulary object navigation in simulation and the real world. The paper reports evaluation metrics including LPIPS, PSNR, SSIM, FVD, FID, BEV IoU, 3D IoU, Chamfer distance, SigLIP similarity, object recognition, occupancy accuracy, and occupancy IoU, and states that 3D-Belief improves 2D and 3D imagination quality and downstream embodied task performance compared to state-of-the-art methods. It also reports improved SAT-Real reasoning, especially on motion- and space-related subsets, when a VLM is augmented with world-model rollouts (Yin et al., 12 May 2026).

Map-space belief prediction reports improvements in occupancy and semantic mIoU in simulation, introduces a $\mathbb{R}^3$3 completion threshold defined as $\mathbb{R}^3$4 of semantic voxels with greater than $\mathbb{R}^3$5 certainty, and shows zero-shot transfer to a UR5 shelf setup. The real-world table reports: Random, 72 correctly found, 52 not found, 11 hallucinated; Ours w/o Pushing, 81 correctly found, 52 not found, 6 hallucinated; Ours, 85 correctly found, 35 not found, 7 hallucinated. The paper’s interpretation is that interactive belief propagation helps reveal objects that remain unseen under non-interactive mapping (Marques et al., 28 Feb 2025).

LGBP reports lower KL divergence from exact marginals, fewer iterations than FOBP, and greater robustness to larger domain sizes and higher weight variance when both lifted factors and lifted messages are used. Its complexity discussion states that if propositional region-graph BP costs $\mathbb{R}^3$6, message-based lifting reduces $\mathbb{R}^3$7 and region-based lifting reduces $\mathbb{R}^3$8 (Smith et al., 2016).

Several limitations recur across these papers. ENBP argues that scalar- and vector-only equivariant architectures cannot by themselves represent anisotropic rank-2 uncertainty, and that single-component messages can be physically misleading (Cheng et al., 4 Jun 2026). 3D-Belief assumes a static world, notes that real-world geometry is not inherently metric because training poses come from SfM, and stresses that imagined regions are hypotheses rather than facts (Yin et al., 12 May 2026). Map-space belief prediction identifies dependence on representative simulation data, segmentation quality, dense voxel-grid scalability, fixed-volume assumptions, a closed-world semantic label set, naïve manipulation action sampling, several-second planning time, and a sim-to-real gap (Marques et al., 28 Feb 2025). LGBP, for its part, remains an approximate inter-region propagation method even though it performs exact lifted inference within regions (Smith et al., 2016).

Across these works, a consistent misconception is rejected: spatial belief is not merely occupancy with uncertainty bars, nor merely visually plausible hallucination, nor merely message reuse by symmetry. The papers instead suggest a broader technical principle: belief masses become substantially more useful when they are explicitly mapped onto the geometry, topology, or region structure over which inference and action must operate.