PROFusion: Iterative Fusion Across Domains

• PROFusion is a multifaceted term that designates distinct approaches for progressive integration in machine learning, RGB-D reconstruction, and quantum information.
• In multimodal learning, Progressive Fusion feeds fused context back into unimodal encoders to recover lost cross-modal information, achieving improvements such as up to 5% MSE reduction and 40% robustness gains.
• In robotic SLAM and quantum memory, PROFusion systems combine learned pose regression with geometric refinement or leverage symmetry-protected fragmentation to deliver robust performance under challenging conditions.

Searching arXiv for “PROFusion” and closely related uses to ground the entry in current literature. Tool call: arxiv_search({"query":"PROFusion OR \"Progressive Fusion\" OR \"ProFusion3D\" OR \"Robust and Accurate Dense Reconstruction via Camera Pose Regression and Optimization\" OR \"Profusion of Symmetry-Protected Qubits\"", "max_results": 10, "sort_by": "relevance"}) PROFusion is not a single universally fixed technical term in the arXiv literature. Instead, it designates several distinct research programs whose common element is usually an idea of progressive or profuse structuring, but whose technical content differs substantially across fields. In multimodal machine learning, Progressive Fusion—also stylized as Pro-Fusion—is an iterative representation-refinement mechanism that feeds fused multimodal context back into earlier unimodal encoders in order to mitigate late-fusion information loss (Shankar et al., 2022). In RGB-D dense reconstruction, PROFusion names a real-time hybrid SLAM system that combines camera pose regression with randomized TSDF-based refinement to remain robust under unstable camera motion (Dong et al., 29 Sep 2025). In nonequilibrium quantum information, PROFusion denotes a proposal for exponentially many symmetry-protected qubits arising from stable ergodicity breaking and topological Hilbert space fragmentation (Iadecola et al., 23 Dec 2025). The term also appears in adjacent but unrelated usages, including ProFusion3D for progressive multi-modal fusion in 3D object detection (Mohan et al., 2024), “profusion” in exoplanet atmospheric spectroscopy (Bello-Arufe et al., 2021), and a “profusion” of classically $1/2$ BPS Wilson loops in $\mathcal N=4$ Chern–Simons–matter theories (Cooke et al., 2015).

1. Terminological status and cross-domain usage

Within machine learning, the earliest direct match is “Progressive Fusion for Multimodal Integration” (Shankar et al., 2022). That work explicitly names the method “Progressive Fusion” and also stylizes or abbreviates it as “Pro-Fusion”. Its central problem is the well-known tension between late fusion, which preserves modality-specific specialization but risks discarding conditionally relevant information before modalities interact, and early fusion, which exposes cross-modal dependencies early but is difficult under heterogeneous encoders and often increases sample complexity (Shankar et al., 2022).

A separate use appears in robotics and SLAM in “PROFusion: Robust and Accurate Dense Reconstruction via Camera Pose Regression and Optimization” (Dong et al., 29 Sep 2025). Here the capitalization PROFusion is the official system name and refers to a hybrid RGB-D tracking-and-fusion pipeline designed for large viewpoint changes, fast motions, sudden shaking, and rapid in-place rotation (Dong et al., 29 Sep 2025).

A third, conceptually unrelated usage appears in “Profusion of Symmetry-Protected Qubits from Stable Ergodicity Breaking” (Iadecola et al., 23 Dec 2025). In that context, PROFusion is a proposal in which a profusion—an exponentially large number—of encoded qubits emerges by combining a discrete global symmetry with topological Hilbert space fragmentation (Iadecola et al., 23 Dec 2025).

The term also enters neighboring literatures through naming analogies rather than shared formalism. ProFusion3D is a LiDAR-camera fusion detector for autonomous driving that performs progressive fusion across Bird’s Eye View and Perspective View at both intermediate-feature and object-query levels (Mohan et al., 2024). In astronomy, “profusion” describes the unusually large atmospheric inventory detected in the ultra-hot Jupiter HAT-P-70 b (Bello-Arufe et al., 2021). In high-energy theory, “profusion” refers to the unexpectedly large set of classically $1/2$ BPS Wilson-loop constructions in certain $3$d $\mathcal N=4$ Chern–Simons–matter theories (Cooke et al., 2015). This suggests that the lexical overlap is broad, whereas the technical content is field-specific.

2. Progressive Fusion in multimodal learning

In multimodal representation learning, Progressive Fusion is formulated by starting from a standard supervised multimodal model

$$\mathcal{F}(x)=P\big(F(G_1(x_1),G_2(x_2),\dots,G_K(x_K))\big),$$

where $G_i$ are unimodal feature generators, $F$ is the fusion operator, and $P$ is the prediction head (Shankar et al., 2022). The motivating concern is the late-fusion failure mode described as “fuse it or lose it”: if a unimodal encoder compresses away information that becomes relevant only after conditioning on another modality, the final fusion layer cannot recover it (Shankar et al., 2022).

The proposed remedy is an iterative representation-refinement scheme in which the fused multimodal representation is projected back into each modality-specific encoder. The augmented model introduces a context vector $c\in\mathbb R^d$ such that $\mathcal N=4$0 recovers the original architecture. In the recurrence form given in the appendix,

$\mathcal N=4$1

$\mathcal N=4$2

$\mathcal N=4$3

where $\mathcal N=4$4 is the backprojection, $\mathcal N=4$5 is an embedding or projection, and $\mathcal N=4$6 is the number of refinement steps (Shankar et al., 2022). The recurrence is over fusion refinement steps, not over input time.

Architecturally, the defining feature is the addition of backprojective / backward connections or skip-back connections from the late fused representation to earlier unimodal branches. The proposal is explicitly model-agnostic: it does not specify a new fusion operator $\mathcal N=4$7, but instead augments existing late-fusion systems by conditioning unimodal feature extraction on prior fused context (Shankar et al., 2022). The paper argues that this restores some of the off-diagonal cross-modal interactions absent from standard late fusion, while retaining the modularity that makes late fusion practical for heterogeneous modalities (Shankar et al., 2022).

Empirically, the method is evaluated on synthetic data, AV-MNIST multimedia classification, CMU-MOSI and CMU-MOSEI sentiment prediction, and financial time-series prediction. The headline result is that Progressive Fusion consistently improves performance, with the strongest reported gains being up to 5% reduction in MSE and about 40% relative robustness improvement on multimodal stock prediction (Shankar et al., 2022). On sentiment benchmarks, the gains are smaller—summarized by the authors as roughly 2% accuracy improvement—which the paper attributes to settings in which text alone already carries much of the predictive signal (Shankar et al., 2022).

3. PROFusion as a hybrid RGB-D dense reconstruction system

In RGB-D SLAM, PROFusion addresses a different failure mode: dense reconstruction under unstable camera motion. The system takes an RGB-D video $\mathcal N=4$8, estimates camera poses

$\mathcal N=4$9

and incrementally reconstructs geometry in a TSDF volume (Dong et al., 29 Sep 2025). The paper’s argument is that classical optimization-based tracking is accurate but brittle under large motions because it requires good initialization, whereas learned pose estimation is robust to large viewpoint changes but not precise enough for dense reconstruction on its own (Dong et al., 29 Sep 2025).

The system therefore combines two stages per frame. First, a camera pose regression network predicts the relative pose between consecutive RGB-D frames. Second, this estimate is used as the initialization for randomized geometric refinement against the accumulated TSDF map. The initialized world pose is

$1/2$0

The network is based on a DUSt3R-style two-branch Vision Transformer. RGB images are patch-embedded into color tokens, while depth is back-projected into metric point clouds and patch-embedded into geometry tokens that are not normalized and are not passed through the same encoder, specifically to preserve metric scale (Dong et al., 29 Sep 2025). Training uses a metric relative-pose loss

$1/2$1

with geodesic angular error on $1/2$2 and Euclidean translation error (Dong et al., 29 Sep 2025).

Refinement is a depth-only randomized search in pose space. Candidate updates are evaluated with a volumetric TSDF-consistency objective

$1/2$3

and the pose is iteratively updated by averaging the improving hypotheses. The search size is adapted according to

$1/2$4

(Dong et al., 29 Sep 2025). The paper stresses that the refinement stage is not point-to-plane ICP and does not use photometric terms.

The reported runtime is real time: pose regression takes < 20 ms, randomized optimization < 10 ms, and the full system runs at > 30 FPS, with total GPU memory remaining below 10 GB in all experiments (Dong et al., 29 Sep 2025). On stable TUM RGB-D sequences, PROFusion remains competitive with global pipelines despite using only single-frame tracking; on unstable-motion benchmarks it is markedly stronger. The strongest quantitative result is on FastCaMo-Synth, where the average ATE-RMSE is 0.7 cm versus 2.6 cm for ROSEFusion on raw data and 1.5 cm versus 2.9 cm under motion blur and depth noise (Dong et al., 29 Sep 2025). A central ablation shows that PR alone is robust but drifts, RO alone is accurate but not robust enough, and PR + RO yields both robustness and dense-reconstruction-grade alignment (Dong et al., 29 Sep 2025).

4. PROFusion in stable ergodicity breaking and quantum memory

In quantum many-body physics, PROFusion refers to a symmetry-enriched fragmentation mechanism rather than a multimodal or reconstruction method. The proposal combines a discrete symmetry with topological Hilbert space fragmentation so as to produce exponentially many encoded qubits that are protected by symmetry and stabilized by the topological obstruction to local sector mixing (Iadecola et al., 23 Dec 2025).

The explicit construction is based on the periodic square-lattice $1/2$5 model with Hamiltonian

$1/2$6

and global

$1/2$7

symmetry generated by

$1/2$8

(Iadecola et al., 23 Dec 2025). In the $1/2$9 limit, the effective Hamiltonian becomes

$3$0

so a site is flippable only if all four of its nearest neighbors are equal (Iadecola et al., 23 Dec 2025). The frozen product states in this limit correspond, in the dual loop picture, to close-packed configurations of noncontractible loops.

The exact number of frozen states is

$3$1

Each frozen state $3$2 has symmetry partners

$3$3

forming a 4-state orbit that encodes two logical qubits. The total number of encoded qubits is therefore

$3$4

(Iadecola et al., 23 Dec 2025). The key robustness claim is that changing the topological sector requires flipping $3$5 qubits along a noncontractible loop, so arbitrary local symmetry-respecting perturbations can only mix sectors in perturbative order scaling with system size (Iadecola et al., 23 Dec 2025).

The work emphasizes that this is not a conventional quantum error-correcting code. Although the construction admits a universal set of transversal logical gates for the paired qubits, the authors explicitly invoke the Eastin–Knill obstruction to explain why this cannot be a full fault-tolerant QEC code (Iadecola et al., 23 Dec 2025). The protection is instead against a restricted class of perturbations: symmetric local perturbations with locality scale $3$6 satisfying

$3$7

Under these conditions, the encoded qubits are stated to be stable for times exponentially long in $3$8 (Iadecola et al., 23 Dec 2025). This suggests a passive many-body quantum-memory paradigm rather than a standard stabilizer-code construction.

5. Related progressive-fusion variants and neighboring uses of “profusion”

The most direct adjacent method name is ProFusion3D, a LiDAR-camera 3D object detector for autonomous driving (Mohan et al., 2024). Its motivation differs from Progressive Fusion (Shankar et al., 2022): the problem is not late-fusion bottlenecks in arbitrary multimodal learning, but the loss of complementary information when fusion is performed in only one view, typically either BEV or PV (Mohan et al., 2024). The architecture therefore projects LiDAR features into PV and camera features into BEV, fuses both views at the intermediate-feature level with an Inter-Intra Fusion block, then refines object queries first separately per view and then jointly. The paper reports 71.1 mAP / 73.6 NDS on nuScenes and 37.7 mAP / 29.1 CDS on Argoverse2, together with robustness under missing-modality conditions, reaching 63.9 mAP with only LiDAR and 38.9 mAP with only cameras on nuScenes (Mohan et al., 2024).

Outside machine learning and quantum information, the word profusion appears descriptively rather than as a stable method name. In exoplanet spectroscopy, “Mining the Ultra-Hot Skies of HAT-P-70b: Detection of a Profusion of Neutral and Ionized Species” reports a rich atmospheric inventory from a single HARPS-N transit, with secure detections of $3$9, $\mathcal N=4$0, $\mathcal N=4$1, $\mathcal N=4$2, $\mathcal N=4$3, $\mathcal N=4$4, $\mathcal N=4$5, $\mathcal N=4$6, and $\mathcal N=4$7, plus tentative $\mathcal N=4$8 and $\mathcal N=4$9 (Bello-Arufe et al., 2021). In that context, “profusion” simply denotes the unusually large number of species identified.

A still more distant usage occurs in “A profusion of $$\mathcal{F}(x)=P\big(F(G_1(x_1),G_2(x_2),\dots,G_K(x_K))\big),$$0 BPS Wilson loops in $$\mathcal{F}(x)=P\big(F(G_1(x_1),G_2(x_2),\dots,G_K(x_K))\big),$$1 Chern-Simons-matter theories”, where the term refers to the unexpectedly large set of classically $$\mathcal{F}(x)=P\big(F(G_1(x_1),G_2(x_2),\dots,G_K(x_K))\big),$$2 BPS Wilson-loop candidates associated with quiver segments and zero-level nodes (Cooke et al., 2015). The authors argue that this abundance is likely classical and that only one linear combination should remain truly BPS once quantum corrections are taken into account (Cooke et al., 2015). These examples underscore that “PROFusion” and “profusion” are semantically flexible labels whose encyclopedia treatment requires disambiguation by discipline.

6. Conceptual contrasts, misconceptions, and significance

A common misconception is to treat PROFusion as a single technical framework. The literature does not support that reading. Progressive Fusion / Pro-Fusion (Shankar et al., 2022), PROFusion for RGB-D dense reconstruction (Dong et al., 29 Sep 2025), and PROFusion for symmetry-protected qubits (Iadecola et al., 23 Dec 2025) are independent constructions with distinct mathematical objects, benchmarks, and claims. Their only shared feature is nominal: each uses “fusion” or “profusion” to signal either iterative integration or multiplicity.

Even within machine learning, conflation is misleading. Progressive Fusion (Shankar et al., 2022) is a model-agnostic iterative representation refinement scheme for multimodal learning, whereas ProFusion3D (Mohan et al., 2024) is a specific camera-LiDAR 3D detector whose novelty lies in progressive multi-view fusion across BEV and PV, together with self-supervised mask-modeling pre-training. PROFusion for RGB-D reconstruction (Dong et al., 29 Sep 2025), by contrast, is not a fusion operator in the same sense at all; it is a tracking-and-reconstruction system whose hybrid design combines a learned relative-pose regressor with geometric optimization.

Another misconception is to read the shared vocabulary as implying shared theoretical commitments. In fact, the technical meanings diverge sharply. In (Shankar et al., 2022), “progressive” denotes repeated refinement via backward connections from fused context to unimodal encoders. In (Dong et al., 29 Sep 2025), robustness emerges from a learned initializer that expands the convergence basin of randomized TSDF alignment. In (Iadecola et al., 23 Dec 2025), “profusion” denotes an exponential scaling of encoded qubits produced by symmetry-related topological fragmentation sectors.

The broader significance of the term is therefore bibliographic rather than doctrinal. It marks a recurring stylistic preference for naming methods that either progressively integrate information or generate a large multiplicity of protected, detected, or classically allowed structures. A plausible implication is that future uses of PROFusion will continue to require local disciplinary qualification—Progressive Fusion in multimodal learning, PROFusion in SLAM, PROFusion in fragmented quantum memories, or ProFusion3D in autonomous-driving perception—because the name by itself does not uniquely determine the underlying method or theory.