Aykut Koç and colleagues have published “Graph Neural Networks Improve Quantized Transformers by Incorporating Global Relationships” in IEEE Transactions on Signal and Information Processing over Networks (January 2026).
Large language models are costly to run, and compressing their parameters to low precision, known as quantization, usually hurts performance. The paper introduces BitTransGNN, which combines quantized transformers with graph neural networks during training and substantially reduces, and often closes, this performance gap while keeping the memory and energy savings. This makes efficient language models more practical on limited hardware.
Authors: Cihan Eralp Kumbasar, Haldun M. Özaktaş, Aykut Koç