Quantized Model Soup Shake-Up: Weight Perturbation for Enhanced Ensemble Diversity
Model soup, averaging the weights of multiple fine-tuned models, delivers ensemble-level accuracy at single-model inference cost, but its success requires both linear mode connectivity (LMC) and sufficient diversity among candidates. We study these two requirements under quantization-aware training (QAT). First, we show analytically and empirically that fine-tuning from a sufficiently converged QAT checkpoint, which we call the QAT anchor state, preserves linear mode connectivity by keeping models within the same loss basin. Second, we identify Ensemble Degeneracy, where the many-to-one mapping of the quantizer collapses independently fine-tuned models into nearly identical quantized representations, eliminating diversity despite intact connectivity. To resolve this, we propose Quantized Model Soup Shake-Up (QMSS), which selectively perturbs low-magnitude weights near quantization bin boundaries to flip their integer assignments, then briefly re-trains each variant via QAT. Because only a small fraction of inherently unstable weights are modified, QMSS induces substantial quantized-domain diversity while preserving LMC. Experiments on CIFAR-100, Tiny-ImageNet, ImageNet, and FSD-Kaggle2018 with both EWGS and LSQ quantizers on CNNs and Vision Transformers show consistent improvements over standard model soup, with gains up to +1.54% Top-1 on CIFAR-100 and +2.09% on CIFAR-100-C, alongside over 6.37× inference speedup on a Jetson Orin Nano at 8/8-bit.