Extreme low-bit quantization-aware training (QAT) is often unstable due to synchronized bin transitions, which can trigger catastrophic loss spikes. While prior works rely on scale-side controls to mitigate this fragility, we demonstrate that restricting pivot movement alone is insufficient. Instead, we identify a critical weight-side state governing low-bit QAT stability: pivot-proximal residence, where latent weights persistently linger near quantization pivots. We formalize this duration as dwell time and show that long dwell times accumulate a large, vulnerable population of weights prone to synchronized mass flips. To resolve this vulnerability, we propose Pivot Gradient Scaling (PGS) as a principled modification to the Straight-Through Estimator (STE) that leaves the forward pass intact and selectively amplifies latent-weight gradients within a localized pivot band. Crucially, we provide a mathematical guarantee proving that this sign-preserving amplification increases pivot-proximal mobility and systematically shortens expected dwell time under both SGD and Adam. Exhaustive empirical evaluations across Llama 3.2 and Qwen3 model families under full-training regimes up to 30 billion tokens demonstrate that PGS structurally stabilizes optimization, attenuates transition bursts, and delivers substantial downstream accuracy gains.

Recent research demonstrates that word embeddings, trained on the human-generated corpus, have strong gender biases in embedding spaces, and these biases can result in the discriminative results from the various downstream tasks. Whereas the previous methods project word embeddings into a linear subspace for debiasing, we introduce a Latent Disentanglement method with a siamese auto-encoder structure with an adapted gradient reversal layer. Our structure enables the separation of the semantic latent information and gender latent information of given word into the disjoint latent dimensions. Afterwards, we introduce a Counterfactual Generation to convert the gender information of words, so the original and the modified embeddings can produce a gender-neutralized word embedding after geometric alignment regularization, without loss of semantic information. From the various quantitative and qualitative debiasing experiments, our method shows to be better than existing debiasing methods in debiasing word embeddings. In addition, Our method shows the ability to preserve semantic information during debiasing by minimizing the semantic information losses for extrinsic NLP downstream tasks.