WorksComputer Science
Topological Deep Learning Challenge 2026: Bridging the Gap
Abstract
This paper describes the 2026 edition of the Topological Deep Learning Challenge: Bridging the Gap, hosted at the second Topology, Algebra, and Geometry in Data Science (TAG-DS) Conference. Where previous editions grew the modelling and data ecosystem of Topological Deep Learning (TDL), this edition set out to place Graph Neural Networks (GNNs) and Topological Neural Networks (TNNs) side by side under a single evaluation pipeline. Participants implemented recent state-of-the-art architectures in TopoBench, with Track 1 covering GNNs and Track 2 covering TNNs. They evaluated these implementations on shared synthetic benchmarks generated with GraphUniverse, which independently controls homophily, average degree and degree-distribution heaviness. The challenge yielded 68 qualifying submissions from 34 teams, spanning the graph, simplicial, cellular, hypergraph and combinatorial domains. For the first time, submissions were also tracked on a public leaderboard, which we summarize and analyze here. Our analysis finds that the differences between the submitted tracks are strongly task-dependent: the mean accuracy gap is modest on community detection, while on triangle counting the Track 1 mean error is 1.85 times the Track 2 mean. Several structurally informed GNNs also outperform the Track 2 mean on triangle counting.
Main organizers, equal contribution. Corresponding email: [email protected]
Introduction
Many physical systems, from molecules to galaxies, are driven by interactions among their components: chemical interactions between atoms, functional connectivity between brain regions, or gravitational forces between stars. Conventionally, these relations are modeled with graphs, allowing Graph Neural Networks (GNNs) (Scarselli et al., 2008) to extract key insights (Kipf and Welling, 2017). However, these systems also involve richer interactions that extend beyond the pairwise relations graphs encode. Topological Neural Networks (TNNs) have emerged to address this limitation by accounting for higher-order interactions (Hajij et al., 2022; Bodnar, 2022; Papillon et al., 2023b), extending message passing to simplicial, cellular, hypergraph and combinatorial domains using concepts from algebraic topology [10, 3].
Despite this shared lineage, the GNN and Topological Deep Learning (TDL) communities have largely operated in parallel. The two fields are often presented as distinct paradigms, yet they are deeply intertwined, and a growing body of work studies the interplay between standard message passing and higher-order representations. Comparisons between GNNs and TNNs remain difficult because published results typically use different codebases, datasets, tasks, and tuning budgets. This leaves a central empirical question: when does higher-order structure improve performance?
Building on the three previous editions of the challenge (Papillon et al., 2023a; Bernárdez et al., 2024; Bernárdez et al., 2026), the Topological Deep Learning Challenge 2026 set the stage for closing this gap. Participants were asked to implement recent state-of-the-art architectures from either literature into TopoBench (Telyatnikov et al., 2025), the open-source platform for developing and benchmarking TDL models, and to evaluate them through a single shared pipeline on synthetic graphs generated with GraphUniverse (Van Langendonck et al., 2026). Because GraphUniverse generates graphs with prescribed structural properties, the resulting benchmark allows us to examine how each architecture responds to changes in homophily, average degree and degree-distribution heaviness under a controlled generative design.
The challenge concluded with 68 valid submissions from 34 teams, submitted by 45 registered participants. Thirty-seven submissions implemented graph architectures in Track 1; thirty-one implemented higher-order architectures in Track 2, distributed across the simplicial, cellular, hypergraph and combinatorial domains. For the first time in the history of the challenge, submissions were additionally evaluated on a common benchmark suite and tracked on a public leaderboard. This paper details the challenge setup and submission requirements, lists all submissions and winners, and presents an analysis of the leaderboard results.
Setup of the Challenge
The challenge1 was hosted at the second Topology, Algebra, and Geometry in Data Science (TAG-DS) Conference and sponsored by Arlequin AI and New Theory. Participants were asked to implement a recent state-of-the-art message-passing model into TopoBench in one of two tracks:
Track 1 — Graph Neural Networks (GNNs). Architectures operating on graph structures, including message-passing networks and graph transformers with structural encodings.
Track 2 — Topological Neural Networks (TNNs). Higher-order message-passing models leveraging rich topological domains, namely simplicial, cellular, hypergraph and combinatorial complexes.
While the tracks were judged separately, the cross-track analysis uses a common evaluation protocol and controlled dataset family.
Guidelines
Participation was free and open to everyone. To enroll, participants had to open a Pull Request (PR) on TopoBench, fill out the Registration Google Form with PR and team information, and tag the PR with the appropriate track label (track-1-gnn or track-2-tnn). Teams were accepted with a maximum of two members, though larger teams were possible upon approval. A single team was permitted to submit multiple PRs, both within and across tracks. The final submission deadline was August 1st, 2026 (AoE).
Submissions were required to implement models already proposed in the literature and to cite the associated publication or pre-print in the PR description. Each PR could contain at most one core model architecture; models with several significant variants were asked to supply a separate configuration file per variant. Participants were encouraged to submit early, both to leave time to resolve integration issues and because, where multiple high-quality submissions covered the same architecture, earlier submissions were given priority.
Consistent with the aims of an open environment for sharing participation, this activity was completely voluntary. However, due to legal restrictions, individuals affiliated with institutions on the Restricted Foreign Research Institutions list were not eligible for reward outcomes, including co-authorship on this white paper.
Submission Requirements
A submission consisted of a Pull Request to TopoBench repository, titled following the format Track: [Track1|Track2]; Team Name: <team name>; Model: <model name>. To be valid, the PR had to pass all integration tests, including linting and formatting. Three components were required of every submission, across tracks:
Backbone. The model implemented as a
torch.nn.Moduleand placed intopobench/nn/backbones/{domain}/{model_name}.py, where{domain}is one of graph, simplicial, cell, hypergraph, combinatorial or non-relational.Hydra configuration. A YAML file at
configs/model/{domain}/{model_name}.yamlspecifying the full TBModel composition: feature encoder, backbone, backbone wrapper and readout.
Results file. The
results.jsonproduced by running the provided evaluation notebook,TopoBench/2026_tdl_challenge/run_evaluation.ipynb, on the submitted model. This file records the model’s performance on the benchmark tasks and available resource measurements, and is what feeds the public leaderboard.
Beyond the backbone, participants were encouraged to exploit TopoBench’s modular architecture by contributing custom feature encoders, readouts, backbone wrappers or losses where their chosen model required them, each inheriting from the corresponding abstract base class and discovered automatically by the framework. This was a deliberate design choice: it enriches the ecosystem with reusable components rather than with monolithic model files.
Testing. All contributed files had to pass the pre-existing unit tests, and any method or class not already covered had to be accompanied by new tests mirroring the source directory structure. Coverage, measured with Codecov, had to match or exceed 93%. Each submission also had to fill in test/pipeline/test_pipeline.py to demonstrate that the full training pipeline runs end-to-end, and the evaluation notebook had to run successfully with the submitted model.
Evaluation Method
To decide the winners in each track, a panel of TopoBench maintainers and collaborators voted on their preferred submissions, and the Condorcet method was used to produce a ranking. Reviewers were asked to base their rankings on the following criteria, which were visible to participants throughout:
Correctness: Does the submission correctly implement the model as described in the literature? Modifications required to respect TopoBench’s computational constraints were permitted.
Code quality: How readable and clean is the implementation? How well does the submission respect the requirements, including unit tests and memory usage?
Benchmark on GraphUniverse datasets: Is the model correctly benchmarked on the provided datasets? Does it produce reasonable results?
Documentation & tests: Do docstrings clearly describe the code, and do they make explicit reference to the original equations in the associated paper and/or the original implementation? Are unit tests robust?
The award criteria assess the quality and reusability of model implementations, not their final benchmark performance. No minimum training performance was required, and the highest leaderboard score did not determine the winner. The leaderboard supports the comparative analysis in § 5.
For the first time, reviewers were permitted to use AI tools to help manage the volume of submissions. In Track 1, an AI reviewer agent was given the same instructions as the human reviewers. It graded each submission against the rubric and flagged the bottom tier of submissions. Track 1 reviewers could use these flags at their discretion. Reviewers were asked to disclose their use of AI in their responses.
Prizes
With thanks to our sponsors, this edition offered cash prizes of $1,000 USD for first place and $400 USD for second place in each track, with Track 1 sponsored by New Theory and Track 2 by Arlequin AI. This year, honorable mentions also carried cash prizes, with $700 USD split across outstanding submissions. In addition, selected teams were invited for research internships at the Geometric Intelligence Lab (University of California, Santa Barbara) and the Intelligent Maintenance and Operations Systems (IMOS) Lab (École Polytechnique Fédérale de Lausanne, Switzerland).
Software Practices
All submitted code had to comply with the challenge’s GitHub Action workflow, passing all tests, linting and formatting checks. To ensure consistency, participants working in the simplicial, cellular or combinatorial domains were asked to use TopoNetX’s (Hajij et al., 2024) classes to manage the corresponding complexes.
Benchmarking Protocol
The defining novelty of this edition is the use of a shared evaluation pipeline and dataset family for the cross-track comparison. This section describes that protocol.
Controlled Synthetic Data
Evaluation datasets were generated with GraphUniverse (Van Langendonck et al., 2026), a framework for sampling synthetic graphs with prescribed structural properties. Three structural axes were varied independently:
Homophily, at three levels: low (0–0.1), mid (0.4–0.6) and high (0.9–1.0).
Average degree, at two levels: low (1–2.5) and high (4–5).
Power-law exponent of the degree distribution, at two levels: low (1.5–2, heavy-tailed) and high (4–5, homogeneous).
The full grid therefore contains configurations. The prescribed evaluation uses three training seeds per configuration, totaling 36 runs per task and 72 per model variant. This controlled design allows us to examine how model performance changes across prescribed structural regimes.
Tasks and Metrics
We consider two tasks, probing community structure and local substructure:
Community detection, a node-classification task scored by accuracy (higher is better). This task is well aligned with the inductive bias of standard message passing: it rewards models that aggregate and smooth information over neighbourhoods.
Triangle counting, a regression task scored by (lower is better), where counts triangles in each test graph’s undirected edge structure after preprocessing. Counting triangles is provably beyond the reach of 1-WL-bounded message-passing networks.
Leaderboard
Results were published on a public leaderboard2. For each structural setting, scores are reported as the mean standard deviation of the in-distribution test metric across three training seeds; summaries across settings include up to 36 runs per task. The analysis uses the submitted result reports, with one variant per PR selected as described in § 5.
Two caveats underscore these results. First, the scores describe the implementations and configurations submitted by participants rather than the best achievable performance of each method. Second, implementations were contributed by different participants with differing familiarity with the underlying papers. While the review process checked for correctness as part of the rubric, residual implementation variance potentially impacts leaderboard performance.
Submissions and Winners
The challenge received 68 valid submissions from 34 teams, submitted by 45 registered participants. Participation was close to evenly split between the two tracks: 37 submissions in Track 1 and 31 in Track 2. Within Track 2, submissions covered all four higher-order domains supported by TopoBench: ten simplicial, six cellular, seven hypergraph and eight combinatorial. Table 1 lists all submitted models, grouped by track and domain, with references to the underlying methods and links to their implementations. Additionally, Table 4 in Appendix A lists the co-authors belonging to each team, as well as the awards each team received.
| Track 1 — GNNs | ||
| Graph | 37 submissions | |
| Model | PR | |
| AdvDIFFormer (Wu et al., 2025) | 399 | |
| BuNN ([2]) | 319 | |
| CoGNN (Finkelshtein et al., 2024) | 369 | |
| Conn-NSD ([4]) | 327 | |
| DNSD ([16]) | 375 | |
| DSNN (Fiorini et al., 2026) | 412 | |
| ESC-GNN (Yan et al., 2024) | 419 | |
| FAGCN ([11]) | 382 | |
| FAGCN ([11]) | 388 | |
| FiGURe (Ekbote et al., 2023) | 360 | |
| GAMMA (Ghazizadeh et al., 2025) | 411 | |
| GATE (Mustafa and Burkholz, 2024) | 355 | |
| GAUGE (Sun et al., 2026) | 396 | |
| GFT (Wang et al., 2024) | 334 | |
| GPR-GNN (Chien et al., 2021) | 352 | |
| GREAD (Choi et al., 2023) | 415 | |
| GREAD (Choi et al., 2023) | 420 | |
| GRIT (Ma et al., 2023) | 370 | |
| GSN (Bouritsas et al., 2023) | 407 | |
| HeroFilter (Zhang et al., 2025a) | 384 | |
| HiD-Net (Li et al., 2024) | 335 | |
| HOD-GNN (Eitan et al., 2026) | 402 | |
| HyMN (Southern et al., 2025) | 421 | |
| ISN (Hernandez Caralt et al., 2026) | 374 | |
| JacobiConv (Wang and Zhang, 2022) | 383 | |
| LINKX (Lim et al., 2021) | 361 | |
| MTGCN (Pei et al., 2024) | 400 | |
| NSD ([15]) | 332 | |
| NSP (Suk et al., 2022) | 326 | |
| NSP + Holonomy (Suk et al., 2022) | 373 | |
| OGFormer (Zhang et al., 2025b) | 378 | |
| PolyFilter GNN (Favard) (Guo and Wei, 2023) | 354 | |
| Polynormer (Deng et al., 2024) | 345 | |
| r-lMPNN (Paolino et al., 2024) | 404 | |
| SGFormer (Wu et al., 2023) | 359 | |
| SheafAN ([5]) | 322 | |
| TopoPolynormer (Deng et al., 2024) | 353 | |
| Track 2 — TNNs | ||
| Simplicial | 10 submissions | |
| Model | PR | |
| Bi-SCNN (Yan and Kuruoglu, 2025) | 387 | |
| COSIMO (Einizade et al., 2025) | 392 | |
| CTNN (Hajij et al., 2025) | 405 | |
| Dir-SNN (Lecha et al., 2025) | 328 | |
| DirSNN (memory-safe, harmonic-aware) (Lecha et al., 2025) | 350 | |
| GSAN ([6]) | 365 | |
| HiGCN (Huang et al., 2024) | 364 | |
| MPSN ([14]) | 376 | |
| MPSN ([14]) | 393 | |
| PolynomialFilterTNN (SNN) (Ebli et al., 2020) | 414 | |
| Cellular | 6 submissions | |
| AirTNN (Fiorellino et al., 2024) | 329 | |
| AirTNN (Fiorellino et al., 2024) | 394 | |
| Cellular Transformer ([1]) | 371 | |
| CIN ([13]) | 363 | |
| DiffLift-SMCN (Franco et al., 2026; Eitan et al., 2025) | 406 | |
| SheafTSP (Tandon et al., 2026) | 410 | |
| Hypergraph | 7 submissions | |
| DPHGNN (Saxena et al., 2024) | 398 | |
| DSHN (Mule et al., 2026) | 408 | |
| EHNN (Kim et al., 2022) | 366 | |
| HyperGCN (Yadati et al., 2019) | 413 | |
| PhenomNN (Wang et al., 2023) | 367 | |
| SheafHyperGNN (Duta et al., 2023) | 321 | |
| WHNN (Duta and Liò, 2025) | 417 | |
| Combinatorial | 8 submissions | |
| CTNN (copresheaf, gated) (Hajij et al., 2025) | 418 | |
| ETNN ([7]) | 320 | |
| ETNN-CoordinatePolicy (structural-LapPE) ([7]) | 391 | |
| ETNN-LapPE ([7]) | 362 | |
| HMC (Hajij et al., 2022) | 416 | |
| SMCN (Eitan et al., 2025) | 397 | |
| TopNets (Verma et al., 2024) | 333 | |
| TopoU-Net (Gaurav et al., 2026) | 372 |
Table 1. All 68 submissions, grouped by track and domain, with models ordered alphabetically within each domain. Citations identify the underlying methods, and PR numbers link to each implementation. Repeated models denote separate submissions. Appendix A lists teams; Appendix B reports scores for the 65 entries in the comparative analysis.
| Team | Co-authors | Track | PR numbers | Award |
| AdamGohain | Adam Gohain | T1 | 359 | |
| CoPreSheafers | Christine Zeh, Lennart Bastian | T2 | 418 | |
| CtrlAltMe | Salwa Khattami | T2 | 397 | |
| DeveshKaushal | Devesh Kaushal | T2 | 392 | |
| E(n)igma | Gaurav Khanal | T2 | 320, 362, 391 | T2 joint 1st (ETNN-CoordinatePolicy) |
| Fernando Espinosa | Fernando David Espinosa Iñiguez | T2 | 410 | |
| Francesco Orsi | Francesco Orsi | T2 | 393, 394 | |
| GiggioIlGriggio | Alessio Comparini | T2 | 376 | |
| Grold | Grold Otieno Mboya | T2 | 350 | |
| howyadoin | Dario Loi, Gabriele Onorato | T1+T2 | 322, 326, 405, 408, 412 | T1 1st (DSNN); HM portfolio |
| Kajolya | Kolya Lettl, Kajal Singla | T1 | 407 | |
| kxdenz | Kaden Zheng | T1 | 419 | |
| LangDiff | Jonas Müller | T1 | 335, 399, 400, 420 | |
| Lettlini | Kolya Lettl, Ghazal Mostofi | T1 | 396 | |
| LoopCounter | Sud Athreya | T2 | 364, 365, 366, 367 | |
| MISN_AY | Mohammad Yusuf, Ankit Kumar | T1+T2 | 382, 383, 387, 388 | |
| One Ring to Lift Them All | Jorge Luiz Franco | T2 | 406 | HM differentiable liftings |
| Oversmooth operators | Gabriel Franco, Yelizaveta Falkouskaya | T1+T2 | 398, 404 | |
| Perelyn Research | Michael Banf, Dominik Filipiak | T1 | 361 | |
| Prasanna28Devadiga | Prasanna Devadiga | T1 | 327 | |
| PushparajD | Pushparaj Devadiga | T1 | 352, 355 | |
| r2 | Anier Velasco Sotomayor | T1+T2 | 329, 354, 360, 414 | |
| Raj Patil | Raj Patil | T1 | 411 | |
| rayquaza | Luigi Petti, Eric Dolores-Cuenca | T2 | 413 | |
| Remy | Stefano Blando | T2 | 417 | |
| s/pairwise/ho | Valentina Sánchez | T2 | 321 | |
| Sheafu | Ankit Grover, Rémi Bourgerie | T1 | 373, 374, 375 | HM supplementary analysis |
| SweetLesson | Aarav Gupta, Riyan Patel | T1 | 345, 353, 402 | T1 joint 2nd (HOD-GNN) |
| TJPaik | Taejin Paik | T1+T2 | 328, 332, 333, 334 | |
| TopoCoop | Marcos Chávez Mesa, Gabriel Chaviano Díaz | T1+T2 | 369, 371 | |
| topolab | Chenguang Xu | T1+T2 | 415, 416 | |
| TripleA | Amirreza Akbari, Alireza Mohammadian | T1+T2 | 363, 370, 372, 378, 384 | T1 joint 2nd (GRIT); T2 joint 1st (TopoU-Net); HM portfolio |
| weeyev | Vihan Tiwari | T1 | 421 | |
| yeahkitory | Hisataka Kajita | T1 | 319 |
Table 4. Teams and their co-authors, in alphabetical order of team name. T1 and T2 denote Track 1 (GNN) and Track 2 (TNN); HM denotes an honorable mention (§ 4). A dash marks a team for which no author information was provided.
Winners and Honorable Mentions
The winners were announced publicly at the second TAG-DS Conference, on social media, and on the official challenge website. We emphasize that regardless of final rankings, the organizers and reviewers were impressed by the quality of the submissions and their contribution to the TopoBench ecosystem. We warmly congratulate all participants.
Track 1 — Graph Neural Networks.
First place ($1,000): DSNN, Team howyadoin (Dario Loi and Gabriele Onorato; PR #412).
Joint second place ($200 each): GRIT, Team TripleA (Amirreza Akbari and Alireza Mohammadian; PR #370), and HOD-GNN, Team SweetLesson (Aarav Gupta and Riyan Patel; PR #402).
Track 2 — Topological Neural Networks. The Condorcet evaluation resulted in a tie, and the Track 2 prize was divided equally between the two co-winners:
Joint first place ($700 each): ETNN-CoordinatePolicy, Team E(n)igma (Gaurav Khanal; PR #391), and TopoU-Net, Team TripleA (Amirreza Akbari and Alireza Mohammadian; PR #372).
Honorable mentions.
Excellent portfolio ($150 each): Teams howyadoin (Dario Loi and Gabriele Onorato) and TripleA (Amirreza Akbari and Alireza Mohammadian), recognised for multiple high-quality contributions across both tracks.
Novel infrastructure with differentiable liftings ($200): DiffLift-SMCN, Team One Ring to Lift Them All (Jorge Luiz Franco; PR #406).
Outstanding supplementary analysis ($200): Neural Sheaf Propagation + Holonomy Measurement, Team Sheafu (Ankit Grover and Rémi Bourgerie; PR #373 with a companion paper Grover and Bourgerie (2026)).
Leaderboard Results and Analysis
This section presents the first cross-track comparison in the history of the challenge. We include 65 entries3: 34 in Track 1 and 31 in Track 2. For each PR, we consider variants of the submitted model with complete benchmark results, select the variant with the highest mean in-distribution community-detection accuracy over the full evaluation grid, and use that same variant for both tasks4. For each task, we first average the three seed scores within each setting, then average the 12 setting means to obtain one score per PR. Track and domain means give each PR equal weight. Table 2 reports the per-track aggregates, and Table 3 breaks them down by topological domain. Complete leaderboards for both tasks are given in Appendix B.
| Track | All | Low | Mid | High | Low | High | |
| Community detection — accuracy () | |||||||
| Track 1 (GNN) | 34 | 0.4241 | 0.3088 | 0.3869 | 0.5765 | 0.3991 | 0.4491 |
| Track 2 (TNN) | 31 | 0.4343 | 0.3061 | 0.4000 | 0.5970 | 0.4099 | 0.4588 |
| Triangle counting — MSE per triangle () | |||||||
| Track 1 (GNN) | 34 | 0.7097 | 0.0887 | 0.6099 | 1.4304 | 0.3579 | 1.0614 |
| Track 2 (TNN) | 31 | 0.3828 | 0.0428 | 0.3251 | 0.7805 | 0.1858 | 0.5798 |
Table 2. Mean in-distribution test score per track, averaged over submissions in the 65-entry analysis cohort (§ 5). Columns marginalise the GraphUniverse configuration grid over homophily and average degree. PRs #319, #327 and #382 are excluded.
| Domain | Community detection () All | Triangle counting () All | |||||||
| Graph (T1) | 34 | 0.4241 | 0.3088 | 0.3869 | 0.5765 | 0.7097 | 0.0887 | 0.6099 | 1.4304 |
| Simplicial | 10 | 0.4411 | 0.3141 | 0.4030 | 0.6062 | 0.2586 | 0.0119 | 0.2184 | 0.5455 |
| Cell | 6 | 0.4356 | 0.2846 | 0.3994 | 0.6227 | 0.3144 | 0.0467 | 0.2652 | 0.6314 |
| Hypergraph | 7 | 0.4451 | 0.3196 | 0.4067 | 0.6090 | 0.6505 | 0.0867 | 0.5253 | 1.3393 |
| Combinatorial | 8 | 0.4155 | 0.3004 | 0.3907 | 0.5555 | 0.3551 | 0.0403 | 0.3281 | 0.6970 |
Table 3. Mean in-distribution test score per topological domain, broken down by homophily regime. Track 1 submissions all operate on the graph domain. The 65-entry analysis cohort is defined in § 5.
Performance Differences Across Tasks
The clearest takeaway from these results is that the mean performance advantage of TNNs over GNNs is task-dependent. For example, it is modest on community detection and more pronounced on triangle counting.
On community detection, Track 2 attains a mean accuracy of 0.4343 against 0.4241 for Track 1, a modest descriptive difference of 1.02 percentage points. The best individual result comes from Track 1: GRIT (PR #370) reaches 0.5207, ahead of the best Track 2 model, HiGCN (PR #364), at 0.4869.
On triangle counting, the picture changes sharply. Track 2 attains a mean error of 0.3828 against 0.7097 for Track 1; the Track 1 mean is 1.85 times the Track 2 mean. The best submission overall is a combinatorial model, CTNN (PR #418), at 0.0024—more than two orders of magnitude below the Track 1 mean. This is consistent with the motivation behind encoding higher-order structure. We note that the track comparison does not isolate the contribution of that structure from other architectural and implementation choices.
Table 3 gives a more detailed view. On triangle counting, simplicial submissions have the lowest mean error (0.2586), followed by cellular (0.3144), combinatorial (0.3551), hypergraph (0.6505) and graph submissions (0.7097).
Structural Information Beyond Track Labels
Going beyond track aggregations, we find that several of the strongest triangle-counting submissions are Track 1 models. For example, TopoPolynormer (PR #353, 0.0451), ESC-GNN (PR #419, 0.0552), GSN (PR #407, 0.0665), GRIT (PR #370, 0.0693) and HOD-GNN (PR #402, 0.0802) all outperform the Track 2 mean, and several outperform the majority of Track 2 submissions. A common property of these architectures is that while they operate on graph-based domains, they inject substructure or structural-encoding information, such as subgraph counts, random-walk encodings, or explicit higher-order descriptors.
These results show that a graph-domain model equipped with structural information can perform competitively on triangle counting. Higher-order domains offer another way to supply structural information, but performance also varies substantially among TNN submissions. Establishing the separate contributions of the lifting, structural encoding and message-passing scheme via controlled ablations is a promising avenue for future work.
Structural Sensitivity
The controlled grid allows us to ask how each family responds to the structural axes. Figure 1 reveals substantial overlap between GNNs and TNNs on community detection. At high homophily, their median accuracies are close (0.6170 and 0.6161, respectively), despite the larger difference between their means. On triangle counting, TNNs achieve lower median error across all three homophily regimes, showing that their advantage is also reflected in the medians. Nevertheless, performance varies considerably within each family, and several GNN submissions achieve errors below the TNN median in every homophily regime.

Figure 1. Submission-score distributions by homophily, complementing Table 2. Each dot is one submission’s mean over four degree–power-law settings and three seeds per setting (34 GNNs; 31 TNNs). Boxes span the interquartile range, with a line marking the median. Whiskers extend to the most extreme observations within 1.5 interquartile ranges below the first quartile or above the third quartile. Diamonds mark arithmetic track means. Axes are shared within each task; triangle-counting error is shown on a logarithmic scale.
Homophily. On community detection, both tracks lose accuracy as homophily drops: from 0.5765 to 0.3088 for Track 1 and from 0.5970 to 0.3061 for Track 2. The Track 2 mean exceeds the Track 1 mean by 2.04 percentage points at high homophily and 1.31 at mid homophily, while Track 1 leads by 0.28 percentage points at low homophily. Thus, the observed Track 2 advantage does not extend to the low-homophily regime on this task. On triangle counting, the direction is reversed: mean error decreases as homophily falls, from 1.4304 to 0.0887 for Track 1 and from 0.7805 to 0.0428 for Track 2.
Average degree and degree distribution. Both tracks perform better on community detection at high average degree than at low (0.4491 versus 0.3991 for Track 1; 0.4588 versus 0.4099 for Track 2), consistent with denser neighborhoods carrying more signal. On triangle counting the pattern inverts, with normalized error rising at high degree and at high homophily in both tracks. Heavy-tailed degree distributions (low ) also yield higher mean normalized error than homogeneous ones in both tracks.
Leaderboard Position and Award Outcome
Because both a leaderboard and a set of awards appear in this paper for the first time, we make their relationship explicit. As described in § 2.3, awards were decided by Condorcet vote on implementation quality, not on benchmark score, and the two rankings differ considerably: the Track 1 winner, DSNN (PR #412), ranks 62nd of 65 on community detection, and the Track 2 co-winner TopoU-Net (PR #372) ranks 59th. The awards recognize implementation quality and reusability; Table 5 and Table 6 describe performance on the two benchmark tasks. Award outcomes and leaderboard positions should therefore be interpreted separately.
| # | Model | Track | Team | Score |
| 1 | Graph Inductive Bias Transformer (GRIT) | T1 · GNN | TripleA | 0.5207 |
| 2 | Filter Bank GNN (FiGURe) | T1 · GNN | r2 | 0.4986 |
| 3 | PolyFilter GNN (Favard) | T1 · GNN | r2 | 0.4977 |
| 4 | HiGCN | T2 · TNN | LoopCounter | 0.4869 |
| 5 | GREAD | T1 · GNN | topolab | 0.4816 |
| 6 | TopoPolynormer | T1 · GNN | SweetLesson | 0.4798 |
| 7 | Deep Neural Sheaf Diffusion (DNSD) | T1 · GNN | Sheafu | 0.4743 |
| 8 | ESC-GNN | T1 · GNN | kxdenz | 0.4743 |
| 9 | MPSN | T2 · TNN | GiggioIlGriggio | 0.4742 |
| 10 | SheafTSP | T2 · TNN | Fernando Espinosa | 0.4735 |
| 11 | SheafHyperGNN | T2 · TNN | s/pairwise/ho | 0.4728 |
| 12 | DiffLift-SMCN | T2 · TNN | One Ring to Lift Them All | 0.4719 |
| 13 | Generalized PageRank GNN | T1 · GNN | PushparajD | 0.4697 |
| 14 | MTGCN | T1 · GNN | LangDiff | 0.4697 |
| 15 | CTNN (copresheaf, gated) | T2 · TNN | CoPreSheafers | 0.4690 |
| 16 | HiD-Net | T1 · GNN | LangDiff | 0.4685 |
| 17 | Cell Isomorphism Network (CIN) | T2 · TNN | TripleA | 0.4681 |
| 18 | JacobiConv | T1 · GNN | MISN AY | 0.4649 |
| 19 | MPNN | T1 · GNN | Oversmooth operators | 0.4648 |
| 20 | DPHGNN | T2 · TNN | Oversmooth operators | 0.4641 |
| 21 | DSHN | T2 · TNN | howyadoin | 0.4633 |
| 22 | HMC | T2 · TNN | topolab | 0.4596 |
| 23 | CoGNN | T1 · GNN | TopoCoop | 0.4585 |
| 24 | HeroFilter | T1 · GNN | TripleA | 0.4574 |
| 25 | OGFormer | T1 · GNN | TripleA | 0.4574 |
| 26 | EHNN | T2 · TNN | LoopCounter | 0.4566 |
| 27 | CTNN | T2 · TNN | howyadoin | 0.4561 |
| 28 | GSAN | T2 · TNN | LoopCounter | 0.4556 |
| 29 | Polynormer | T1 · GNN | SweetLesson | 0.4550 |
| 30 | ETNN-LapPE | T2 · TNN | E(n)igma | 0.4545 |
| 31 | ETNN-CoordinatePolicy (structural-LapPE) | T2 · TNN | E(n)igma | 0.4543 |
| 32 | ETNN | T2 · TNN | E(n)igma | 0.4534 |
| 33 | GATE | T1 · GNN | PushparajD | 0.4457 |
| 34 | GFT | T1 · GNN | TJPaik | 0.4451 |
| 35 | WHNN | T2 · TNN | Remy | 0.4435 |
| 36 | DirSNN (Memory-Safe, Harmonic-Aware) | T2 · TNN | Grold | 0.4414 |
| 37 | HyMN | T1 · GNN | weeyev | 0.4408 |
| 38 | Cellular Transformer | T2 · TNN | TopoCoop | 0.4400 |
| 39 | HOD-GNN | T1 · GNN | SweetLesson | 0.4395 |
| 40 | MPSN | T2 · TNN | Francesco Orsi | 0.4387 |
| 41 | SGFormer | T1 · GNN | AdamGohain | 0.4375 |
| 42 | PolynomialFilterTNN (SNN) | T2 · TNN | r2 | 0.4368 |
| 43 | COSIMO | T2 · TNN | DeveshKaushal | 0.4333 |
| # | Model | Track | Team | Score |
| 44 | PhenomNN | T2 · TNN | LoopCounter | 0.4308 |
| 45 | GAMMA | T1 · GNN | Raj Patil | 0.4257 |
| 46 | GAUGE (Riemannian Graph Foundation Model) | T1 · GNN | Lettlini | 0.4219 |
| 47 | Neural Sheaf Diffusion (NSD) | T1 · GNN | TJPaik | 0.4113 |
| 48 | Bi-SCNN | T2 · TNN | MISN_AY | 0.4057 |
| 49 | AdvDIFFormer | T1 · GNN | LangDiff | 0.4023 |
| 50 | AirTNN | T2 · TNN | Francesco Orsi | 0.4007 |
| 51 | GREAD | T1 · GNN | LangDiff | 0.3937 |
| 52 | SheafAN | T1 · GNN | howyadoin | 0.3887 |
| 53 | Neural Sheaf Propagation | T1 · GNN | howyadoin | 0.3852 |
| 54 | HyperGCN | T2 · TNN | rayquaza | 0.3847 |
| 55 | Directed Simplicial Neural Networks (Dir-SNN) | T2 · TNN | TJPaik | 0.3821 |
| 56 | Neural Sheaf Propagation + Holonomy Measurement | T1 · GNN | Sheafu | 0.3817 |
| 57 | FAGCN | T1 · GNN | MISN_AY | 0.3813 |
| 58 | SMCN | T2 · TNN | CtrlAltMe | 0.3697 |
| 59 | TopoU-Net | T2 · TNN | TripleA | 0.3666 |
| 60 | LINKX | T1 · GNN | Perelyn Research | 0.3628 |
| 61 | AirTNN | T2 · TNN | r2 | 0.3594 |
| 62 | DSNN | T1 · GNN | howyadoin | 0.3081 |
| 63 | TopNets | T2 · TNN | TJPaik | 0.2970 |
| 64 | Identity Sheaf Network (ISN) | T1 · GNN | Sheafu | 0.2755 |
| 65 | GSN | T1 · GNN | Kajolya | 0.0795 |
Table 5. Community-detection leaderboard for the 65-entry analysis cohort (accuracy, higher is better; § 5). Ranks use full-precision means.
| # | Model | Track | Team | Score ↓ |
| 1 | CTNN (copresheaf, gated) | T2 · TNN | CoPreSheafers | 0.0024 |
| 2 | SheafTSP | T2 · TNN | Fernando Espinosa | 0.0108 |
| 3 | MPSN | T2 · TNN | GiggioIlGriggio | 0.0287 |
| 4 | MPSN | T2 · TNN | Francesco Orsi | 0.0333 |
| 5 | TopoPolynormer | T1 · GNN | SweetLesson | 0.0451 |
| 6 | COSIMO | T2 · TNN | DeveshKaushal | 0.0504 |
| 7 | ESC-GNN | T1 · GNN | kxdenz | 0.0552 |
| 8 | GSN | T1 · GNN | Kajolya | 0.0665 |
| 9 | Graph Inductive Bias Transformer (GRIT) | T1 · GNN | TripleA | 0.0693 |
| 10 | HOD-GNN | T1 · GNN | SweetLesson | 0.0802 |
| 11 | ETNN-LapPE | T2 · TNN | E(n)igma | 0.1147 |
| 12 | TopNets | T2 · TNN | TJPaik | 0.1192 |
| 13 | DiffLift-SMCN | T2 · TNN | One Ring to Lift Them All | 0.1201 |
| 14 | ETNN-CoordinatePolicy (structural-LapPE) | T2 · TNN | E(n)igma | 0.1211 |
| 15 | ETNN | T2 · TNN | E(n)igma | 0.1213 |
| 16 | -MPNN | T1 · GNN | Oversmooth operators | 0.1382 |
| 17 | HyMN | T1 · GNN | weeyev | 0.1497 |
| 18 | Bi-SCNN | T2 · TNN | MISN AY | 0.1498 |
| 19 | DirSNN (Memory-Safe, Harmonic-Aware) | T2 · TNN | Grold | 0.1543 |
| 20 | EHNN | T2 · TNN | LoopCounter | 0.1672 |
| 21 | Directed Simplicial Neural Networks (Dir-SNN) | T2 · TNN | TJPaik | 0.1716 |
| 22 | GSAN | T2 · TNN | LoopCounter | 0.2570 |
| 23 | Cell Isomorphism Network (CIN) | T2 · TNN | TripleA | 0.2638 |
| 24 | DPHGNN | T2 · TNN | Oversmooth operators | 0.4021 |
| 25 | PolynomialFilterTNN (SNN) | T2 · TNN | r2 | 0.4197 |
| 26 | AirTNN | T2 · TNN | Francesco Orsi | 0.4538 |
| 27 | HMC | T2 · TNN | topolab | 0.4843 |
| 28 | GFT | T1 · GNN | TJPaik | 0.4911 |
| 29 | Cellular Transformer | T2 · TNN | TopoCoop | 0.4982 |
| 30 | CTNN | T2 · TNN | howyadoin | 0.5174 |
| 31 | Neural Sheaf Diffusion (NSD) | T1 · GNN | TJPaik | 0.5378 |
| 32 | AirTNN | T2 · TNN | r2 | 0.5401 |
| 33 | WHNN | T2 · TNN | Remy | 0.5492 |
| 34 | Identity Sheaf Network (ISN) | T1 · GNN | Sheafu | 0.6044 |
| 35 | SheafHyperGNN | T2 · TNN | s/pairwise/ho | 0.6742 |
| 36 | HiGCN | T2 · TNN | LoopCounter | 0.8038 |
| 37 | SMCN | T2 · TNN | CtrlAltMe | 0.8219 |
| 38 | Neural Sheaf Propagation | T1 · GNN | howyadoin | 0.8292 |
| 39 | Filter Bank GNN (FiGURe) | T1 · GNN | r2 | 0.8396 |
| 40 | PolyFilter GNN (Favard) | T1 · GNN | r2 | 0.8425 |
| 41 | Deep Neural Sheaf Diffusion (DNSD) | T1 · GNN | Sheafu | 0.8658 |
| 42 | Polynormer | T1 · GNN | SweetLesson | 0.8721 |
| 43 | HiD-Net | T1 · GNN | LangDiff | 0.8753 |
| 44 | GAUGE (Riemannian Graph Foundation Model) | T1 · GNN | Lettlini | 0.8758 |
| 45 | GREAD | T1 · GNN | topolab | 0.8773 |
| 46 | DSHN | T2 · TNN | howyadoin | 0.9073 |
| 47 | DSNN | T1 · GNN | howyadoin | 0.9099 |
| 48 | LINKX | T1 · GNN | Perelyn Research | 0.9132 |
| # | Model | Track | Team | Score ↓ |
| 49 | Neural Sheaf Propagation + Holonomy Measurement | T1 · GNN | Sheafu | 0.9158 |
| 50 | MTGCN | T1 · GNN | LangDiff | 0.9160 |
| 51 | GATE | T1 · GNN | PushparajD | 0.9223 |
| 52 | PhenomNN | T2 · TNN | LoopCounter | 0.9243 |
| 53 | CoGNN | T1 · GNN | TopoCoop | 0.9286 |
| 54 | HyperGCN | T2 · TNN | rayquaza | 0.9289 |
| 55 | GAMMA | T1 · GNN | Raj Patil | 0.9301 |
| 56 | GREAD | T1 · GNN | LangDiff | 0.9315 |
| 57 | Generalized PageRank GNN | T1 · GNN | PushparajD | 0.9325 |
| 58 | AdvDIFFormer | T1 · GNN | LangDiff | 0.9342 |
| 59 | SheafAN | T1 · GNN | howyadoin | 0.9345 |
| 60 | JacobiConv | T1 · GNN | MISN_AY | 0.9378 |
| 61 | HeroFilter | T1 · GNN | TripleA | 0.9470 |
| 62 | OGFormer | T1 · GNN | TripleA | 0.9509 |
| 63 | SGFormer | T1 · GNN | AdamGohain | 0.9658 |
| 64 | FAGCN | T1 · GNN | MISN_AY | 1.0438 |
| 65 | TopoU-Net | T2 · TNN | TripleA | 1.0560 |
Table 6. Triangle-counting leaderboard for the 65-entry analysis cohort (MSE per triangle, lower is better; § 5). Ranks use full-precision means.
Discussion & Concluding Remarks
This edition set out to answer two questions: how specific graph properties affect classical GNNs versus higher-order TNNs, and under what data regimes TNNs provide capabilities that standard GNNs do not. The results above provide a first, qualified answer to both.
On the first question, the community-detection gap narrows as homophily falls, with a small reversal in favour of GNNs at low homophily; a higher-order domain alone therefore does not establish robustness in that regime. On the second question, the submitted Track 2 cohort has lower mean error on triangle counting, while its mean accuracy advantage on community detection is small. The strong triangle-counting results of several Track 1 models suggest that both higher-order architectures and structurally informed graph models merit consideration for this task. Extending this observation to other motifs or tasks requires further evaluation.
We highlight the finding that substructure-aware GNNs close much of the gap on triangle counting as it points to a possible reframing of the comparison. Rather than a strict dichotomy between two model families, the results motivate studying the structural information supplied to models in both tracks. Both communities therefore face the same practical question: how to give networks access to structure beyond edges.
Several limitations bound these conclusions. The benchmark comprises two tasks on one family of synthetic graphs; performance on real-world data with correlated structural properties may differ. The results concern the submitted configurations, and their relative performance may change under different tuning budgets. Selecting variants by reported test accuracy can favour PRs with more variants and make community-detection performance estimates optimistic. Implementation quality, while reviewed, inevitably varies. Moreover, the aggregate statistics we report summarize a heterogeneous population of architectures, so track-level means should be interpreted as descriptions of the submissions rather than defined properties of the underlying model classes. Extending the benchmark suite with additional tasks and re-running it with per-model tuning are the natural next steps. Both are straightforward within TopoBench.
The challenge yielded 68 valid submissions from 34 teams, enriching TopoBench with implementations from both the GNN and TDL literatures. It also produced the first cross-track comparison in this challenge series under a shared benchmarking protocol, supported by a public leaderboard. We hope this infrastructure hosted in TopoBench will support research within the communities, facilitate comparisons between models, and encourage the exchange of methods and ideas between them.
Acknowledgments
We thank Arlequin AI and New Theory for sponsoring this edition of the challenge, the Geometric Intelligence Lab (UC Santa Barbara) and the IMOS Lab (EPFL) for supporting the internship opportunities, and the TAG-DS organizers for hosting the challenge. We are grateful to all TopoBench maintainers and collaborators who served on the review panel, and above all to the participants, whose contributions constitute the substance of this paper. N. M., G. B., and M. P. acknowledge partial support from NSF grant 2602079 and the Noyce foundation.
Teams and Co-authors
Table 4 associates each participating team with the individual co-authors of this paper who belong to it, and lists the pull requests and awards of each team. Participants who submitted under more than one team appear once per team.
Full Leaderboards
Tables 5 and 6 give the rankings of the 65 entries included in the comparative analysis, with variant selection and exclusions as described in § 5. Scores are means over the 12 GraphUniverse settings and three training seeds per setting; rankings use unrounded scores.
References
- [1]Rubén Ballester, Pablo Hernández-García, Mathilde Papillon, Claudio Battiloro, Nina Miolane, Tolga Birdal, Carles Casacuberta, Sergio Escalera, and Mustafa Hajji. Attending to Topological Spaces: The Cellular Transformer. arXiv preprint arXiv:2405.14094, 2024. URL https://arxiv.org/abs/2405.14094.
- [2]Jacob Bamberger, Federico Barbero, Xiaowen Dong, and Michael M. Bronstein. Bundle Neural Network for message diffusion on graphs. In International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=scI9307PLG.
- [3]Sergio Barbarossa and Stefania Sardellitti. Topological signal processing over simplicial complexes. IEEE Transactions on Signal Processing, 68:2992–3007, 2020.DOI
- [4]Federico Barbero, Cristian Bodnar, Haitz Sáez de Ocáriz Borde, Michael Bronstein, Petar Veličković, and Pietro Lió. Sheaf Neural Networks with Connection Laplacians. In Proceedings of Topological, Algebraic, and Geometric Learning Workshops 2022, volume 196 of Proceedings of Machine Learning Research, pages 28–36. PMLR, 2022a. URL https://proceedings.mlr.press/v196/barbero22a.html.arxiv.org/abs/2206.08702
- [5]Federico Barbero, Cristian Bodnar, Haitz Sáez de Ocáriz Borde, and Pietro Lió. Sheaf Attention Networks. NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations, 2022b. URL https://openreview.net/forum?id=LIDvgVjpkZr.
- [6]Claudio Battiloro, Lucia Testa, Lorenzo Giusti, Stefania Sardellitti, Paolo Di Lorenzo, and Sergio Barbarossa. Generalized Simplicial Attention Neural Networks. IEEE Transactions on Signal and Information Processing over Networks, 10:833–850, 2024. doi: 10.1109/TSIPN.2024.3485473. URL https://ieeexplore.ieee.org/document/10735150/.DOI
- [7]Claudio Battiloro, Ege Karaismailoğlu, Mauricio Tec, George Dasoulas, Michelle Audirac, and Francesca Dominici. E(n) Equivariant Topological Neural Networks. In International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=Ax3uIiEBVR.arxiv.org/abs/2405.15429
- [10]Christian Bick, Elizabeth Gross, Heather A Harrington, and Michael T Schaub. What are higher-order networks? SIAM Review, 65(3):686–731, 2023.DOI
- [11]Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond Low-frequency Information in Graph Convolutional Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):3950–3957, 2021. doi: 10.1609/aaai.v35i5.16514. URL https://ojs.aaai.org/index.php/AAAI/article/view/16514.DOI
- [13]Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Liò, Guido F. Montúfar, and Michael Bronstein. Weisfeiler and Lehman Go Cellular: CW Networks. In Advances in Neural Information Processing Systems, volume 34, pages 2625–2640, 2021a. URL https://proceedings.neurips.cc/paper/2021/hash/157792e4abb490f99dbd738483e0d2d4-Abstract.html.
- [14]Cristian Bodnar, Fabrizio Frasca, Yuguang Wang, Nina Otter, Guido F. Montúfar, Pietro Liò, and Michael Bronstein. Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 1026–1037. PMLR, 2021b. URL https://proceedings.mlr.press/v139/bodnar21a.html.
- [15]Cristian Bodnar, Francesco Di Giovanni, Benjamin Chamberlain, Pietro Liò, and Michael Bronstein. Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs. In Advances in Neural Information Processing Systems, volume 35, pages 18527–18541, 2022. doi: 10.52202/068431-1346. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/75c45fca2aa416ada062b26cc4fb7641-Abstract-Conference.html.DOI
- [16]Rémi Bourgerie, Sārūnas Girdzijauskas, and Viktoria Fodor. Deep Neural Sheaf Diffusion. arXiv preprint arXiv:2605.19021, 2026. URL https://arxiv.org/abs/2605.19021.