No Triangulation Without Representation: Generalization in Topological Deep Learning

Abstract: Despite an ever-increasing interest in topological deep learning models that target higher-order datasets, there is no consensus on how to evaluate such models. This is exacerbated by the fact that topological objects permit operations, such as structural refinements, that are not appropriate for graph data. In this work, we extend MANTRA, a benchmark dataset containing manifold triangulations, to a larger class of manifolds with more diverse homeomorphism types. We show that, unlike prior claims, both graph neural networks (GNNs) and higher-order message passing (HOMP) methods can saturate the benchmark. However, we find that this is contingent on the right representation and feature assignment, emphasizing their importance in baseline models. We thus provide a novel evaluation protocol based on representational diversity and triangulation refinement. Surprisingly, we find no indication that existing models are capable of generalizing beyond the combinatorial structure of the data. This points towards a research gap in developing models that understand topological structure independent of scale. Our work thus provides the necessary scaffolding to evaluate future models and enable the development of topology-aware inductive biases.

• The paper demonstrates that state-of-the-art TDL models rely on combinatorial artifacts rather than intrinsic topological features.
• It introduces an augmented MANTRA dataset using refined triangulation techniques for rigorous evaluation of topological generalization.
• Empirical results show severe performance decline under triangulation refinements, highlighting the need for topology-aware deep learning architectures.

Generalization in Topological Deep Learning: Analysis and Implications

Introduction

The paper "No Triangulation Without Representation: Generalization in Topological Deep Learning" (2605.06467) critically examines the landscape of topological deep learning (TDL), with a focus on the capacity of current neural architectures to genuinely capture topological properties—rather than superficial combinatorial features—in higher-order datasets. The authors introduce a methodology for evaluating TDL models using manifold triangulations, particularly the extended MANTRA dataset, and systematically probe model generalization via principled data augmentation schemes. The empirical findings challenge prevalent assumptions regarding architectural expressivity in the TDL domain, suggesting a deep methodological gap in the field.

Topological Data Representations and Learning Targets

The manuscript underscores the necessity of distinguishing between the combinatorial and topological aspects of manifold representations. Combinatorial structures, such as simplicial complexes, dual graphs, and Hasse diagrams, can encode the same underlying manifold with disparate granularities. The paper formalizes this distinction by leveraging theoretical invariants: Euler characteristics and orientability, which together provide complete classification information for 2D manifolds.

Central to the study is the expansion and refinement of the MANTRA dataset, which originally suffered from severe label imbalance and incomplete coverage of manifold types. By systematically applying Pachner moves (local triangulation-preserving modifications) and connected sum operations, the authors augment the data to balance classes and increase representational diversity without altering underlying manifold topologies. This augmentation enables stress-testing of model generalization across triangulation refinements.

Methodology: Data Augmentation and Evaluation Framework

The authors establish a rigorous benchmark protocol for TDL evaluation by:

• Extending the MANTRA dataset to cover manifold triangulations in dimensions 2 and 3, ensuring all homeomorphism types are represented and balanced.
• Developing topological data augmentation tools—especially via Pachner moves and connected sums—to diversify the dataset while preserving topological invariance.
• Creating combinatorially out-of-distribution test sets through subdivision schemes (stellar and barycentric), enabling models to be evaluated on incrementally refined triangulations.
• Employing deduplication and isomorphism checks (incidence graph hashing, $f$-vector validation, Weisfeiler-Lehman graph hashing) to ensure the augmented datasets avoid leakage or redundancy.

This evaluation protocol provides a smoke-test benchmark analogous to MNIST for topology-aware learning, enabling fine-grained differentiation between combinatorial and topological generalization capacity.

Model Architectures, Representations, and Encodings

The experimental study benchmarks GNNs (GCN, Residual Gated GCN, Graphormer) and higher-order message passing (HOMP) models (SCCNN, CWN) across multiple representations (simplicial complexes, Hasse diagrams, dual graphs, 1-skeletons) and encoding strategies (random features, degree, RWPE, moment curve). The choice of representation and encoding strongly modulates performance, with findings that contradict previous claims:

• Contradictory finding: Standard GNNs, when provided with appropriate representations and positional encodings, saturate the benchmark on both 2D and 3D manifold classification. This directly refutes prior assertions of their inadequacy relative to HOMP-based models.
• Attention-based architectures (Graphormer) perform on par with CWN in 3D settings given optimal features and representations, indicating that graph representations can indeed capture complex manifold interactions.
• Model computational complexities are analyzed, revealing that higher-order models often become infeasible on large or refined complexes, while GNNs scale efficiently for practical scenarios.

The study demonstrates that predictive success is not inherently tied to model class; instead, representational and encoding choices are critical.

Generalization Across Triangulation Refinements

A pivotal contribution of the paper is the analysis of topological versus combinatorial generalization through out-of-distribution subdivision evaluation. Key results include:

• All models, regardless of architecture, exhibit drastic performance degradation when tested on refined triangulations (stellar subdivisions, barycentric subdivisions), converging to chance-level accuracy even after modest refinements.
• This failure mode exposes the reliance of current methods on combinatorial artifacts rather than intrinsic topological structure. Models do not maintain predictive performance when only the combinatorial structure changes—despite the topology being preserved.
• Theoretical invariants (such as Euler characteristic), computed by deterministic heuristics, outperform neural models in refined settings, further emphasizing the lack of topology-aware learning.

These results reveal that existing TDL models—both GNN and HOMP—do not genuinely learn topological properties. Strong numerical evidence underscores the need for fundamentally new architectures and evaluation methodologies.

Discussion and Implications for Future AI Research

The findings necessitate a critical reassessment of the expressivity claims in TDL. Practical implications include:

• The recommendation to adopt the augmented MANTRA dataset and proposed evaluation protocols (refined subdivisions, balanced classes) as a minimum standard for TDL benchmarking.
• The inadequacy of ad-hoc graph-lifting techniques and small synthetic datasets in representing the challenges pertinent to topology-sensitive domains (physics, drug discovery, materials science).
• The observed dependence on combinatorial structure calls for the integration of topology-aware inductive biases—potentially leveraging explicit computation of topological invariants or novel neural mechanisms capable of abstraction across scales.
• Future architectural development should explicitly decouple representation from combinatorial artifacts, focusing on generalization over homeomorphism rather than isomorphism.

The study predicts that meaningful advances in topology-aware learning will require new hybrid frameworks, possibly combining symbolic and geometric reasoning, with robust evaluations that track generalization across refinement and representation variation.

Conclusion

The paper presents an authoritative empirical and conceptual critique of current TDL models, demonstrating that high performance in standard benchmarks often reflects combinatorial memorization rather than genuine topology learning. By establishing rigorous augmentation and evaluation protocols and providing strong negative results, the work highlights a critical research gap and sets foundational standards for future TDL development. The implications extend to broader AI modeling, emphasizing the importance of theoretical invariance, principled benchmarking, and architectural innovation for structured relational domains.