Towards Graph Foundation Models for Dynamics in Complex Networked Systems: Lessons from Super-Spreader Identification in Multilayer Networks
Michał Czuba, Mateusz Stolarski, Adam Piróg, Piotr Bielak, Piotr Bródka
Abstract
Network dynamics - including spreading, influence maximisation, and epidemic modelling - remain largely confined to the transductive paradigm, where models are trained on a single network and cannot be reused on unseen graphs without retraining. We argue that inductive cross-network generalisation is a necessary prerequisite for Graph Foundation Models (GFMs) in this domain and propose four design properties towards this goal. As a proof of concept, ts-net (TopSpreadersNetwork), trained solely on synthetic multilayer networks (MLNs), demonstrates zero-shot generalisation to real-world MLNs of varying size and layer count, outperforming classical heuristics and transductive baselines on three of four metrics. Based on ts-net's performance, we further outline five open challenges towards building GFMs for network dynamics: scale, many-layer generalisation, self-supervised pretraining, cross-task transfer, and node-attribute integration.
AI Impact Assessments
(1 models)Scientific Impact Assessment
1. Core Contribution
This workshop paper makes two intertwined contributions: (1) a conceptual argument that inductive cross-network generalization is a necessary but unmet prerequisite for Graph Foundation Models (GFMs) in the domain of network dynamics (spreading, influence maximization, epidemic modeling), and (2) empirical evidence via ts-net, a GNN-based model trained on synthetic multilayer networks (MLNs) that achieves zero-shot generalization to unseen real-world MLNs for super-spreader identification.
The paper proposes four design desiderata (D1–D4): relationship-agnostic encoding, size-agnostic inference, corpus-based training, and topology-focused input. These are framed as a "candidate minimal set" for GFMs in network dynamics. The conceptual framing is clear: the transductive status quo—where models are trained per-graph and cannot transfer—is explicitly contrasted with the inductive paradigm that GFMs demand.
The problem being addressed is genuine. Network dynamics methods (heuristics, deep-im, PIANO, GBIM) are indeed predominantly transductive, and the multilayer setting amplifies the challenge due to variable layer counts and types. The paper correctly identifies this as a gap.
2. Methodological Rigor
The empirical evidence is drawn primarily from prior work (Czuba et al., 2025), with this paper serving as a position/framing wrapper. The architecture (~458K parameters) uses GAT+GIN encoders with a WiseAverage cross-layer aggregation and neighborhood sampling. Training on 200+ synthetic MLNs (Erdős–Rényi and preferential attachment) with ICM simulations is straightforward but limited in scale.
Strengths in evaluation:
Weaknesses:
3. Potential Impact
The paper's primary value is agenda-setting rather than methodological. It articulates a clear research direction—GFMs for network dynamics—and provides initial evidence of feasibility. The five open challenges (scale, many-layer generalization, self-supervised pretraining, cross-task transfer, node-attribute integration) constitute a reasonable roadmap.
Practical impact is currently limited: ts-net addresses only super-spreader ranking under ICM in MLNs, a relatively niche task. However, the conceptual framework could inspire work on transferable models for influence maximization, epidemic forecasting, and source detection more broadly. The observation that deep-im fails on a 61K-actor network due to memory exhaustion while ts-net handles it via neighborhood sampling is practically relevant.
The multilayer network focus is both a strength (an underserved area) and a limitation (smaller community compared to single-layer graph learning).
4. Timeliness & Relevance
The paper is timely. GFMs are a hot topic (OFA, GraphGPT, etc.), and the observation that network dynamics has been left behind is valid. The workshop venue (GFM @ ICML 2026) is appropriate. However, the gap between "position paper with proof of concept" and "actual GFM for network dynamics" remains substantial, and the paper is transparent about this.
5. Strengths & Limitations
Key Strengths:
Key Limitations:
Additional Observations
The paper's honesty about its limitations and its clear articulation of open challenges are commendable for a workshop paper. The framing of D4 (topology-focused input) as a *feature* rather than a limitation is intellectually interesting but may limit applicability to dynamics that are not purely structure-driven. The reproducibility potential seems reasonable given the description of the synthetic data generation process and architecture, though the reliance on the companion preprint for full details is a drawback.
Generated Jun 9, 2026
Comparison History (24)
Paper 1 addresses a fundamental limitation in graph machine learning (transductive learning) and proposes a roadmap for Graph Foundation Models in complex network dynamics. Its focus on zero-shot generalization across diverse networks has broad, cross-disciplinary implications for epidemiology, social sciences, and complex systems, offering a paradigm shift with higher potential for widespread theoretical and foundational impact compared to the domain-specific robotic manipulation improvements in Paper 2.
Paper 1 addresses the timely and high-impact topic of Graph Foundation Models (GFMs) for network dynamics, proposing a framework for inductive cross-network generalization. This connects to the rapidly growing foundation model paradigm and has broad applications across epidemiology, social networks, and complex systems. While Paper 2 makes solid theoretical contributions to online convex optimization with capacity constraints, it addresses a more niche problem with narrower audience. Paper 1's vision for GFMs in network dynamics, combined with demonstrated zero-shot generalization, positions it for broader interdisciplinary impact and future research directions.
Paper 2 addresses a critical bottleneck in scientific machine learning by effectively modeling multi-scale, non-linear PDEs. Its combination of adaptive multi-scale operators with physics-informed learning provides a rigorous, highly applicable framework for physical sciences and engineering. While Paper 1 introduces an interesting conceptual step toward graph foundation models, it remains primarily a proof-of-concept on synthetic data. Paper 2's methodological rigor, open-source availability, and immediate applicability to complex dynamical systems give it a higher potential for broad and lasting scientific impact.
Paper 2 is likely to have higher scientific impact because it targets a timely, high-visibility direction—graph foundation models and inductive, cross-network generalization for dynamics—potentially affecting multiple application areas (epidemics, influence, multilayer systems) and communities (network science, ML, public health). Even as a proof-of-concept plus agenda paper, it can shape research directions via its design principles and open challenges. Paper 1 is methodologically rigorous and novel for local hypergraph diffusion, but its impact is more specialized to seeded clustering/submodular hypergraph optimization.
Paper 2 has higher potential impact due to broader relevance: inductive, cross-network generalisation for dynamics in complex networks spans epidemiology, information diffusion, infrastructure, and social systems. It proposes general design principles for graph foundation models, demonstrates zero-shot transfer from synthetic to real multilayer networks, and articulates open challenges that can steer a research agenda. Paper 1 is timely and technically solid but is more specialized to LSBO with tabular in-context surrogates and molecular VAEs, yielding narrower cross-field reach.
Paper 1 has higher potential impact due to its broader methodological novelty and cross-domain relevance: it targets inductive, zero-shot generalisation for network dynamics—an underdeveloped capability central to emerging Graph Foundation Models—and demonstrates transfer from synthetic to real multilayer networks. This could influence multiple fields (epidemiology, information diffusion, infrastructure, social systems) and motivates a clear roadmap of open challenges. Paper 2 is strong and clinically relevant, but its contribution is more application-specific (ADNI-based AD progression) and less likely to generalize broadly beyond healthcare longitudinal modeling.
Paper 2 likely has higher impact due to timeliness and broad applicability: improving RL post-training for LLMs is immediately relevant across many deployed systems, and a more stable/efficient trust-region method can propagate widely in industry and academia. Methodologically, it proposes a concrete algorithmic change (smooth divergence regularizer) with extensive empirical validation across scales/settings, suggesting strong rigor and reproducibility. Paper 1 is novel in pushing inductive, zero-shot generalization for network dynamics and could be impactful long-term, but it appears more conceptual/proof-of-concept with narrower near-term adoption compared to LLM RL.
GPT-Micro presents a more complete and validated framework with demonstrated quantitative advantages (70% data reduction, 400X time reduction) for a well-defined, high-impact problem in manufacturing. It integrates multiple innovations—LLM-driven model discovery, thermodynamic compliance, and sparse data utilization—into a novel paradigm with clear practical applications. Paper 2 is more of a proof-of-concept position paper outlining design properties and open challenges for future Graph Foundation Models, with narrower validation (only super-spreader identification) and less mature contributions. Paper 1's broader applicability to manufacturing and stronger empirical validation suggest higher near-term scientific impact.
Paper 1 introduces a paradigm shift towards Graph Foundation Models for network dynamics, offering zero-shot generalization that spans multiple disciplines like epidemiology and social networks. While Paper 2 presents a rigorous and important solution for power grid security, Paper 1's broader scope, alignment with highly impactful AI trends, and potential for widespread cross-disciplinary application give it a higher estimated scientific impact.
Paper 2 targets a broad, timely push toward graph foundation models with clear real-world relevance (epidemics, influence, spreading) and emphasizes inductive, zero-shot cross-network generalization—an impactful capability across domains. It proposes design principles, demonstrates a proof-of-concept model trained on synthetic data transferring to real multilayer networks, and outlines an agenda of open challenges, which can catalyze follow-on work. Paper 1 is novel and rigorous within deep learning theory/diagnostics, but its applications are narrower and primarily interpretive/diagnostic rather than enabling cross-domain deployment.