Community Takes on the Future of Graph Learning
On drifting goals, fresh frontiers, and a bittersweet lesson we are slowly learning
This post was jointly written by the “GLOW organizers” (Fabrizio Frasca, Caterina Graziani, Veronica Lachi, Sohir Maskey, Luis Müller, Maximilian Thiessen, Pascal Welke), along with inputs from Maya Bechler-Speicher, Billy J. Franks, Michael Galkin, Shenyang Huang, Ningyuan (Teresa) Huang, Chaitanya Joshi and Bastian Rieck. This post is, either way, an inherently collaborative effort as it elaborates on emerging community discussions on “The Future of Graph Learning Research”.1
Graph Learning2 has undeniably become a prominent area of research in Machine Learning in recent years. Every four months, a large number of papers are submitted to prominent conferences like ICLR, ICML, or NeurIPS on its behalf. With LoG3, the field even has its own conference now, and we have also witnessed the first industrial and scientific breakthroughs powered by Graph Neural Networks (GNNs) (we’ll get to those soon). What is also undeniable, however, is the presence of this subtle feeling lingering within our community, a feeling which distills into questions such as: “Are we losing relevance?”, “Can we keep up with LLMs, agents, and Gen-AI?” and, most importantly: “What is Graph Learning actually useful for, again?”
In this post, among other things, we will answer these questions. Well… sort of. More concretely, we will present and elaborate on community thoughts in regard to the present and future of Graph Learning research. Everything hereafter is based on discussions held during our GLOW event in February 2025. GLOW (Graph Learning On Wednesdays)4 is a monthly online reading group where we invite researchers to present and discuss their work. Tingled by the aforementioned questions, our session was, however, somewhat atypical. Instead of focusing on a specific paper, we got together with many (~50) Graph Learning folks to simply sit down (in front of our Zoom cameras), have tea (optional), and talk about the state of our field: its problems, its opportunities, and what could come next. For the next few minutes, we hope that you, dear reader, will be at the very least entertained, and unlikely but with nonzero probability inspired by the world of learning on graphs.
Before confronting the here and now, we started the discussion by asking attendees to look back on what the field has achieved.
What can we be proud of?
A first observation we should be making about GNNs is: they actually work, are currently used—in some form—and/or will contribute to new products soon. To start with, we recall how GNNs made an impact in improving Google Maps’ traffic predictor5, in powering recommender systems at Pinterest6, and in developing new antibiotics7. Chaitanya Joshi8 had in mind even more recent examples. Most prominently, the Nobel Prize in chemistry was awarded partially for protein design, with one of the breakthroughs being a system called ProteinMPNN—yes, the trailing part of the acronym stands for Message Passing Neural Network9, the most popular type of GNNs. Another exciting area is molecular simulations, where GNN-driven models learn how atoms interact, as well as their dynamics. Chaitanya believes that we are primed to see breakthroughs in the so-called “Universal Machine Learning Interatomic Potentials” this year (2025). Last, but not least, GNNs are being used every day across pharma and biotech companies, where the hope is that they will contribute to the design of new medicines and biotechnology already in the coming years.
We next asked and discussed which recent idea or development in Graph Learning we found the most insightful or useful, or which research contribution we appreciated the most as a community.
Relevant findings and ideas
Bastian Rieck10 pointed out an interesting recent work that uncovers the tendency of GNNs to always leverage the graph structure for predictions, even when this is totally uninformative. This is the finding of “Graph Neural Networks Use Graphs When They Shouldn't”11 (previously presented at GLOW). He thinks this is a particularly important contribution for our community, and that it sheds light on a crucial takeaway: graph modeling is perhaps at least as important as the graph architectures themselves. Put differently, just modeling a problem as some graph is not necessarily a guarantee for success.
Another open problem is the lack of zero- and few-shot capabilities in GNNs. Such capabilities are at the core of foundation models for text and images12, typically requiring general-purpose representations arising from self-supervised pretraining. To date, no analogous pretraining strategy for Graph Learning has successfully emerged. Billy J. Franks13 highlighted the Graph Positional and Structural Encoder (GPSE)14 as a promising first step towards learning general-purpose graph representations. The underpinning idea is to learn to reconstruct positional and structural encodings in a self-supervised manner; he believes this could serve as a viable objective for large-scale pretraining, potentially paving the way for a foundation model that generalizes across graph-based tasks.
Finally, Ningyuan (Teresa) Huang15 shared an interesting view—ideas, concepts, and approaches developed within the realm of Graph Learning have often contributed to sparking interesting research in adjacent fields. For instance, techniques from GNN expressivity have been utilized for analyzing general transformers: the parallelism between GNNs and the distributed computing model LOCAL16 inspired new results on the equivalence of transformers and Massively Parallel Computation (MPC)17. Moreover, graph reasoning tasks such as connectivity, cycle counts, and shortest paths are now extensively used in understanding the representational power of transformers1819 20, a line of research motivated by the rise of reasoning LLMs.
Reminding ourselves of relevant ideas and findings emerging in Graph Learning is important, yet it shall not distract us from the fact that somehow we are not quite up there in relevance and glamour with the newest reasoning or foundation models. We then asked our attendees how they feel about the current state of Graph Learning and what they believe are the most important research directions and challenges in our field right now.
Graph Learning in the era of LLMs
To start with, Michael Galkin21 shared an interesting view on the current state of Graph Learning research. He believes one could compare it with the state of Reinforcement Learning (RL) around 2018–2022, i.e., post-AlphaZero22 and pre-RL from Human Feedback (RLHF)23. After the successes of AlphaZero, the research community applied RL to almost all closed environments reachable at that time. It celebrated the compelling results but struggled to move beyond game environments. Many felt the field was stagnating and switched to adjacent fields, particularly in light of major setbacks like “Uber selling its robotic vehicle research” in 202024 and “OpenAI disbanding its RL research group” in 202125, and the rather lukewarm progress of RL in robotics (a traditional stronghold of RL). However, the landscape changed radically in 2022 after demonstrations that RLHF was an integral component of ChatGPT and the alignment of LLMs with user preferences. Soon after, RL became the main driver of reasoning models (all the rage in 2025) that solve math olympiad problems, coding competitions, and complex benchmarks that were out of reach for previous-generation LLMs. RL became more popular again and is arguably one of the most promising approaches to advance frontier AI models! Now, it would be reasonable to argue that, despite initial successes, Graph Learning has not found its RLHF moment yet, giving the impression of stagnation similar to RL in 2021. The space of problems tackled by modern Graph Learning may seem to be miles away from modern GenAI and LLM juggernauts. However, at the same time, it is not hard to believe there are plenty of real-world problems where Graph Learning could be extremely impactful, perhaps in the near future—we discuss some of these examples further below in the post.
These thoughts spontaneously led us to discuss whether Graph Learning can intersect with leading research directions in AI. An intriguing perspective, raised by Chaitanya Joshi and Shenyang Huang26, challenges the dominant assumption that sequences are the optimal representation for all data modalities, including language and images. While LLMs and vision transformers have demonstrated that sequential processing can be highly effective, they inherently impose a linear structure on data that may not naturally be sequential. Graphs—capable of capturing richer relational structures—could serve as a more fundamental representation of these modalities. For example, sentences in natural language have an underlying dependency graph, and visual scenes contain objects with structured relationships that are not naturally captured in a sequence of patches. If this viewpoint gained traction, it would not be hard to imagine a paradigm shift where graph-based representations take a more central role in foundation model research, rather than being relegated to more “niche” applications.27
Future venues for Graph Learning research …
Beyond intersecting with current prominent AI research, several other key areas emerged as promising directions for future progress. Billy J. Franks emphasized the importance of anomaly detection, where graph-based methods could be crucial for identifying fraud, cybersecurity threats or, more generally, rare events in complex networks. Meanwhile Shenyang Huang pointed to the development of graph-based foundation models, which could bring us closer to graph representations that generalize well and work across diverse tasks—similar to what LLMs have achieved for text. Bastian Rieck highlighted applications in graph theory itself, suggesting that learning-based approaches could provide new insights into combinatorial problems. While he acknowledges the inherent hardness of such a challenge, his belief is that Graph Learning can assist mathematicians in finding new connections or, in the best case, even resolve open conjectures or prove new results about certain important graph families. Last, he claims, Graph Learning models could play a pivotal role in the time-critical realm of climate-sciences, thanks to their focus on predicting global properties from local information. They could, for instance, support assessments on how the global climate is impacted by (causal) interventions aiming to change local weather patterns, e.g., those implemented through societal changes or some technical means. This is surely another challenging task, but even minor improvements over coarse simulations would be highly beneficial.
… and obstacles ahead
Pursuing these new research frontiers is, however, undoubtedly challenging. We asked our attendees to elaborate on practical roadblocks they see ahead for our community to address.
One of the most widely agreed-upon points was the need to move beyond chasing numbers on artificial benchmarks like ZINC and focus, instead, on real-world applications where Graph Learning can make a tangible impact. While benchmark datasets have played an important role in driving progress, many in the community now recognize that true breakthroughs will come from tackling practical, high-impact problems rather than optimizing performance on toy tasks not representative of real-world problems. This is echoed in the positions of Chaitanya and Bastian, who also put an interesting accent on a problematic “benchmarking culture”. Our community seems to use benchmarks with extreme zeal without necessarily understanding their limitations, and without providing structured information on the data provenance and curation process.28
From a more practical perspective, Maya Bechler-Speicher29 pointed out how the effective application of GNNs often requires intricate expertise. With such a diverse ecosystem of GNN architectures, where each variant addresses specific limitations, practitioners frequently resort to popular models like GCN and GIN, which may deliver suboptimal performance. Interestingly, simpler approaches like combining feature engineering with decision trees can sometimes yield comparable outcomes to vanilla GNNs with lower technical complexity. This often leads to the perception that “GNNs do not work.” In reality, they would—they may just require a higher level of expertise and a tailored approach to unlock their full potential. Going forward, it would be important both to collaborate with domain experts and to design general-purpose methods addressing multiple limitations simultaneously. The aforementioned development of foundation models could also be beneficial in this sense, as practitioners could leverage their outputs for downstream tasks without needing to grapple with the myriad of challenges currently faced to unlock the full power of GNNs.
A piece of advice for early researchers
What emerges from these discussions is, arguably, a somewhat convoluted picture of how a brisk and lively field like Graph Learning is transitioning into a new phase. After exciting years of successes, rapid developments, and high expectations, our community now seeks to find its place within a fast-paced and turbulent research landscape. With our interactions, together we have highlighted past and recent achievements, touched on current directions intersecting with the latest AI trends, and pointed to promising new paths, all while acknowledging practical challenges ahead. At this point, what better way to conclude than by asking Bastian—one of the senior participants in the discussion—what advice he would offer to a new graduate student just beginning their Ph.D. journey in Graph Learning?
We believe the response we received is both reassuring and encouraging. Bastian’s message for new students is not to be intimidated by the current obsession with scaling and large language models. He thinks scale has its place, but that it will only get you so far in scenarios where data and compute are not in abundance. His view is that there is still profound value in understanding the foundations and building efficient models, not just because of a perceived sense of “elegance”, but also because, practically speaking, energy is not free, and efficiency will always be a competitive factor when performances are on-par. According to him, if on the one hand we should heed Sutton’s “Bitter Lesson”30 as a warning against too much fine-tuning to specific tasks, on the other hand creativity and ingenuity still matter, and Graph Learning has still room to grow. His personal advice? To dig deep into the foundations of the fields that one finds interesting to look for cracks therein—there is also value in identifying erroneous assumptions. He hence recommends not to be dismayed when entering a field, but to rather focus on the process of science as such, and, in the words of Francis Bacon: “Read not to contradict and confute, nor to believe and take for granted, nor to find talk and discourse, but to weigh and consider.” A fitting reminder, perhaps, not just for early Ph.D. students, but for our community as a whole.
Authors are entered in alphabetical order by surname.
With “Graph Learning” we broadly refer to machine learning on graph-structured data (and not to the specific inference of structure from unstructured observations).
Chaitanya is a PhD student at the Department of Computer Science and Technology, University of Cambridge, supervised by Prof. Pietro Liò.
Message Passing Neural Networks were introduced as an overarching architectural family in “Neural Message Passing for Quantum Chemistry”. They are amongst the most popular GNN variants nowadays.
Bastian holds the position of Full Professor of Machine Learning at the University of Fribourg in Switzerland, where he leads the AIDOS Lab (Artificial Intelligence for Data-Oriented Science).
Billy is a last year PhD student at the Rheinland-Pfälzische Technische Universität (RPTU), Germany. He is a member of the Machine Learning Group led by Prof. Dr. Marius Kloft.
Sanford et al., “Understanding Transformer Reasoning Capabilities via Graph Algorithms”, NeurIPS 2024.
Merrill and Sabharwal. “The Expressive Power of Transformers with Chain of Thought”, ICLR 2024.
Sanford et al., “Transformers, parallel computation, and logarithmic depth”, ICML 2024.
Abbe et al., “How far can transformers reason? the globality barrier and inductive scratchpad”, NeurIPS 2024.
Silver et al., “Mastering the game of Go without human knowledge”, Nature 2017.
Christiano et al., “Deep reinforcement learning from human preferences”, NeurIPS 2017.
Andy is a final year Ph.D. student at Mila and McGill University, supervised by Professors Reihaneh Rabbany and Guillaume Rabusseau.