Kolmogorov-Arnold Networks Match MLP Performance With 40% Fewer Samples

The researchers behind arXiv preprint 2610.10627 tested how much data Kolmogorov-Arnold networks (KAN) need in reinforcement learning. The answer: 40% less than a multilayer perceptron (MLP) when the goal is comparable quality. During training, KAN's relative gain reached 50%. The experiments ran on the Feynman dataset and the Gymnasium benchmark, and reward noise did not eat into the advantage.

Why samples are the main currency in RL

Deep reinforcement learning has long outperformed classical control methods. In real systems, though, it runs into the cost of experience. Every sample means minutes or hours of robot time, mechanical wear, an occupied test bench, and the risk of breakage. Simulators are cheaper, but the gap between simulation and reality still has to be closed with live data. That is why a method that reaches the target result with fewer interactions with the environment is worth more than one that merely raises peak reward.

KAN was proposed in 2024. The idea rests on the Kolmogorov-Arnold theorem: instead of fixed activation functions at the nodes, the network learns one-dimensional functions on the edges, usually as splines. Previously, two advantages were claimed for such networks: fewer parameters for the same task and better interpretability, especially where a physical relationship needs to be recovered. The new work adds a third axis: data consumption.

What was measured and how

The authors compared KAN and MLP using learning curves. The first testbed is Feynman, a set of tasks that involve recovering physical relationships. The second is Gymnasium, the standard set of RL environments. They also tested robustness to noise separately: they distorted rewards at different levels and checked whether the gap persisted.

Key results from the abstract:

• comparable quality is reached with 40% fewer samples;
• relative improvement during training reaches up to 50%;
• the advantage holds across different levels of reward noise;
• benchmarks: Feynman and Gymnasium;
• the code is open, in the GitHub repository DerKevinRiehl/neurips26_kan_training (the name hints at a NeurIPS 2026 submission).

What remains unclear

The abstract does not say which RL algorithms and which environments were included in the sample, and without that it is hard to judge how well the result transfers to other tasks. The phrase "up to 50%" describes a peak at certain stages of training, not an average gain. This is a first-version preprint and has not been peer reviewed.

There is also a compute caveat. A KAN with splines on its edges is typically more expensive than an MLP in time per training step. Saving samples is not the same as saving GPU hours, and it is worth checking how the authors account for this in the paper. Finally, Feynman tasks are built on smooth physical relationships, which is territory where KANs have historically been strong. For learning from pixels or from highly discontinuous dynamics, the transfer of the gain needs to be verified separately.

What teams should do in practice

Teams where experience is expensive, such as in robotics, equipment control, and industrial processes, should try KAN as a replacement for MLP in the policy or critic on their own task. The code is open, so the comparison will take days, not months.

The math should be done in money: the cost of one sample against the cost of training compute. If a sample from a real test bench costs orders of magnitude more than a GPU hour, even a slow KAN pays off. If the environment is simulated almost for free, the advantage is doubtful. An additional bonus is interpretability: the learned splines can be visualized and checked against the physics of the system, which is useful when certifying control algorithms. The result does not directly apply to LLMs; this is a story about compact control models.