A Preliminary Investigation of Quantum-Assisted Ternary Weight Optimization for BitNet Neural Networks: Proof-of-Concept on Real IBM Quantum Hardware using a Hybrid Classical-Quantum Pipeline

We present a preliminary investigation of quantum-assisted training for BitNet neural networks, which constrain all weights to ternary values -1, 0, +1 for memory efficiency. Classical gradient descent gets trapped in saddle points due to vanishing gradients through the sign () quantization function. We propose Quantum-Assisted Critical Weight Search (QACWS): gradient magnitude analysis identifies the 16 hardest weights for classical training; a QAOA-inspired quantum circuit on real IBM Torino hardware (133 qubits) searches exclusively those weights (1 qubit = 1 real weight) ; classical fine-tuning polishes the result. On a 16-bit XOR benchmark with a 272-weight BitNet model, the hybrid pipeline achieved 59. 0% test accuracy, surpassing the classical plateau of 56. 8% by +2. 2 percentage points. We document a complete five-version experimental trajectory on IBM Torino quantum hardware revealing key failure modes and engineering solutions. All IBM Quantum job IDs are provided for full reproducibility. Source code: github. com/sh1vam-03/bitnetquantum