The paper landed on April 30, 2024. It did not immediately take over the developer conversation. That is worth understanding before evaluating what is happening in 2026.
Liu et al.’s “KAN: Kolmogorov-Arnold Networks” appeared on arXiv with a co-author list that signaled serious institutional backing: Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y. Hou, and Max Tegmark, drawing from MIT, Caltech, and affiliated programs. Tegmark’s involvement alone was a marker, he co-founded the Future of Life Institute and leads MIT’s Machine Intelligence group. This was not a preprint from a grad student. It was a considered research proposal from people with standing in both theoretical and applied ML.
The paper received attention when it published. It did not receive sustained developer adoption. That gap between intellectual interest and production uptake is exactly what makes the 2026 trending event worth analyzing.
What the Paper Actually Claims
The core claim is architectural, and it is confirmed directly in the abstract.
Standard multi-layer perceptrons do two things: they apply fixed activation functions at each neuron, and they learn weights on the connections between neurons. Those learned weights are linear. The activation functions, ReLU, sigmoid, GELU, are fixed at design time. You choose the activation function. The network learns the weights.
KANs invert that arrangement. As the abstract states directly: “KANs have no linear weights at all, every weight parameter is replaced by a univariate function parametrized as a spline.” The activation functions move from nodes to edges, and they become learnable. The network does not choose from a fixed set of activations applied uniformly. It learns the shape of each activation function on each edge, independently, as a spline.
This is a meaningful structural departure. The inspiration is the Kolmogorov-Arnold representation theorem, a mathematical result stating that any multivariate continuous function can be represented as a finite composition of univariate functions and addition. MLPs draw theoretical justification from the universal approximation theorem, which guarantees they can approximate any continuous function given enough neurons. KANs draw from a different theorem with different structural implications, the composition is explicit in the architecture, not just theoretically possible.
The paper’s abstract confirms that this design produces a model that outperforms MLPs in accuracy on the tested tasks. That is a specific, bounded claim. It is not a claim that KANs are better than MLPs in general, or better than Transformers, or ready to replace the backbone of production AI systems. The tested tasks matter, and they were chosen to highlight where the architecture’s properties are most advantageous.
Where KANs Win (and the Boundaries of Those Wins)
The paper’s experimental results focus on two task types: Feynman benchmark datasets and partial differential equation solving. These are scientific and mathematical reasoning tasks, not language generation or multimodal understanding.
Per the paper’s experimental results, KANs outperformed MLPs on Feynman benchmark datasets and in PDE-solving tasks. The interpretability case rests on what that outperformance looks like under inspection: because each edge activation function is a learnable spline, researchers can sometimes read the learned function shapes and recognize mathematical forms. The paper’s authors describe KANs as enabling interpretable scientific modeling, with potential to help researchers identify mathematical structure in physical phenomena.
That interpretability claim is plausible and consistent with the paper’s scope. It is also the claim that requires the most careful handling. It does not mean KANs produce human-readable explanations automatically. It means that the architecture makes certain kinds of post-hoc analysis tractable in ways that MLP weight matrices typically are not. The paper presents this as a research direction and provides examples; it does not prove general interpretability across all task types. Evaluate the full paper text before building workflows that depend on this property.
The paper also reports faster neural scaling laws compared to MLPs in function-fitting experiments, per the authors’ published results. This claim did not appear in the available abstract excerpt and comes from the paper’s body. It is plausible, the spline parametrization may produce more efficient use of parameters on structured mathematical tasks, but it has not been independently replicated at scale by third parties at the time of this writing. It is a result to track, not a production specification to rely on.
What Changed Between 2024 and 2026
Three things shifted the adoption curve.
First, the pykan library matured. Open-source research implementations at time of publication are often proof-of-concept quality, runnable by researchers, not ready for practitioners who need stability, documentation, and integration tooling. The library received five arXiv revisions through February 2025 and ongoing development since. The v5 revision in February 2025 represents substantial refinement from the original April 2024 submission. What developers found harder to use in mid-2024 is more accessible in 2026.
Second, the surrounding context shifted. The benchmark cost conversation has accelerated in a way that makes architectural efficiency arguments more compelling. A model family that achieves comparable or better accuracy with fewer parameters on structured tasks is interesting in the abstract; it is practically interesting when compute costs and evaluation overhead are rising. Scientific AI teams, in drug discovery, materials science, climate modeling, have also grown more sophisticated in their infrastructure requirements, creating a larger addressable audience for an architecture that was always aimed at that segment.
Third, developers simply found it. Viral adoption on developer platforms often reflects a threshold phenomenon: a library reaches a usability level where word-of-mouth becomes self-sustaining. The reported 2,400+ stars in 48 hours on the pykan repository is consistent with that pattern. It is a social signal, not a performance benchmark, but social signals in open-source communities are often leading indicators of sustained adoption.
Epoch AI’s notable models database, updated as of April 29, 2026, does not specifically list KAN in the available content excerpt. This is expected: Epoch tracks deployable frontier commercial systems, and KAN is a research architecture released as an open-source library. Its absence from that database is not a mark against it. It simply means the independent evaluation infrastructure that covers commercial frontier model releases has not been applied to KAN. Third-party reproduction studies exist in the academic literature but have not been consolidated into the kind of standardized comparative evaluation that Epoch provides for commercial models.
A Comparison Frame for Architecture Decisions
The structural differences between KANs and MLPs translate into different tradeoff profiles.
KANs place learnable activation functions on edges, parametrized as splines. MLPs place fixed activation functions on nodes with learned linear weights. In practice, this means KANs offer more expressive per-edge functions, potentially capturing mathematical relationships more efficiently on structured tasks, but they also carry higher per-parameter computational cost during training and inference than equivalent-depth MLPs. The interpretability advantage is real but bounded: it applies most cleanly to tasks where the learned functions can be matched to known mathematical forms. Language modeling, where the learned representations are distributed and high-dimensional, does not present the same interpretability foothold.
For comparison: the Transformer architecture that dominates current LLM development makes different structural choices entirely, attention mechanisms, positional encodings, residual connections, none of which have direct analogues in the KAN framework. KAN is not a Transformer competitor. It is an MLP alternative for a specific class of problems where the MLP’s fixed activation functions are a structural limitation.
What Practitioners Should Actually Do
The evaluation guidance depends entirely on what you are building.
If your work sits in scientific AI, physics simulation, drug discovery, climate modeling, mathematical function learning, PDE-solving workflows, put pykan on your evaluation stack now. The evidence basis for KAN’s advantages in those task categories is direct and comes from credible researchers. Read the full arXiv paper (arXiv:2404.19756), run the library against your specific benchmark tasks, and assess the interpretability properties against your actual interpretability requirements. The paper’s Feynman benchmark results are a starting point, not a guarantee.
If you are building general-purpose language infrastructure, LLMs, RAG systems, agentic pipelines, the current evidence does not support KAN as a primary architecture choice. The performance advantages are demonstrated on mathematical tasks. The training and inference overhead of learnable spline activations on language-scale datasets has not been resolved in available research. The Transformer-derived architectures that underpin current frontier models have years of optimization, tooling, and production hardening behind them. That gap does not close quickly.
The 2026 adoption signal is real. It reflects practitioners catching up to two years of library development and a research paper that had more staying power than its initial reception suggested. What remains open is whether that adoption translates to production results at scale, and whether the interpretability and scaling law claims hold up under independent replication in domains beyond the paper’s original experimental scope.
Read the paper. Test the library. Match the architecture to the problem class. That is the honest evaluation framework the current evidence supports.