Introducing Neumathway: Exploring Neural-Symbolic AI for Mathematical Reasoning

Developer BhinnekaAI released Neumathway, an experimental open-source neural-symbolic model on Hugging Face, trained on 8.27 million samples with 2.47 million parameters to handle mathematical expressions through structured reasoning rather than text prediction. The model targets arithmetic, linear equations, systems of equations, differentiation, integration, limits and matrix operations, combining neural networks with symbolic computation. BhinnekaAI described Neumathway as an ongoing experiment with limitations, shared to gather community feedback on neural-symbolic AI and small specialized models.

Hi everyone I’d like to share Neumathway , an experimental open-source AI project I’m developing to explore a different approach to mathematical reasoning. Unlike conventional language models that generate answers through text prediction, Neumathway explores a neural-symbolic approach , combining neural networks with symbolic mathematical computation to work with mathematical expressions and problems. 8.27 million training samples used to train the model. 2.47 million parameters , making it a relatively compact model. A neural-symbolic architecture designed to combine learned representations with structured mathematical reasoning. Support for multiple mathematical domains, including arithmetic, linear equations, systems of equations, differentiation, integration, limits, and matrix operations. The goal is to explore whether a relatively small, specialized model can perform mathematical reasoning more reliably by combining neural learning with symbolic computation, rather than relying on text generation alone. Large language models have made impressive progress, but mathematical reasoning still presents interesting challenges. I wanted to explore an alternative architecture that treats mathematical problems as structured expressions rather than ordinary text. Neumathway is an ongoing experiment, not a claim that these challenges have been completely solved. It has limitations, and I’m continuing to investigate its capabilities, reliability, and performance. I’m sharing Neumathway with the Hugging Face community to get feedback, discuss the architecture, and learn from others working on neural-symbolic AI, mathematical reasoning, and small specialized models. If you’re interested, feel free to explore the model and share your thoughts Model: BhinnekaAI/neumathwork-v1 · Hugging Face https://huggingface.co/BhinnekaAI/neumathwork-v1 Neumathway is an experimental open-source project, and contributions and feedback are welcome. Thanks for reading