Why Does Straight-Through Work? A Systematic Ablation of Gradient Estimators for Ternary Quantization-Aware Training

A systematic evaluation of eight alternative gradient estimation techniques for training neural networks with ternary {-1, 0, +1} weights. Tested alternatives include DQT stochastic rounding, CAGE curvature-aware correction, QZO zeroth-order optimization, two HESTIA variants, Tequila deadzone reactivation, frozen-weight alpha-only learning, and Integer Gradient Accumulation (IGA). Results demonstrate that the Straight-Through Estimator (STE) remains optimal for ternary quantization-aware training, with IGA as the only marginally viable alternative. Three recurring failure modes are identified: gradient variance explosion, error magnitude mismatch, and training regime incompatibility.