stability

Definition

Not diverging, spiking or varying wildly between runs. At scale this is a concrete engineering concern rather than a theoretical nicety: loss spikes, overflow in reduced precision and seed-to-seed variance can end a training run outright, which is why normalisation choices, learning-rate schedules and fp16/bf16 decisions are all discussed under this heading. Stability is a separate axis from final accuracy.

Explainers using this term

• Paper Walkthrough: Designing Qwen3.8-Next — Accuracy, Efficiency and Stability as One ProblemOn the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability