LLM Uncovers Power-Law Scaling in Small Spiking Neural Networks

A now-withdrawn arXiv paper from January 2026 (revised in June) investigated how classification accuracy scales in small-scale spiking neural networks based on Leaky Integrate-and-Fire (LIF) neurons. The team—Zhengdi Zhang, Cong Han, and Wenjun Xia—systematically varied neuron count, stimulus nodes, and number of categories, then used a large language model (LLM) to help discover the underlying mathematical relationships. Traditional linear and polynomial fitting methods were compared head-to-head against the LLM's ability to propose plausible functional forms.

Key result: classification accuracy follows a power-law scaling primarily determined by the number of categories, while neuron count and stimulus nodes have only minor effects. Notably, the LLM was able to suggest concise and accurate mathematical descriptions beyond predefined equation templates. The paper was withdrawn to improve academic writing and the model, but the findings highlight how LLMs can assist in AI-aided scientific discovery—specifically for uncovering scaling laws in brain-inspired computational systems.