QF2: Quick Firing Model Weight Updates

We propose Quick Firing (QF2) Learning, a novel biologically inspired framework for knowledge consolidation in neural networks. In contrast to conventional deep learning, which relies on gradient backpropagation and complex optimization, QF2 enables direct, one-shot synaptic weight updates driven by firing-based rules, without the need for gradient descent or matrix inversion. This approach closely mimics fundamental mechanisms of biological learning, including engram cell formation and Hebbian plasticity, making it a more brain-like solution for artificial intelligence. We mathematically prove the effectiveness of the QF2 update rule and demonstrate its practical utility through extensive experiments on large language models. QF2 achieves rapid, continual learning and robust knowledge retention, with minimal computational cost and no catastrophic forgetting. Furthermore, the framework is fully open-sourced, providing both code and models for easy reproducibility. Our results suggest that QF2 not only advances the frontiers of brain-inspired AI but also offers new directions for neuroscience and brain–machine interface research.