In long pretraining, what pattern most strongly suggests excessive weight decay?

Too much weight decay is an over-regularization problem. The model stays numerically stable, but it cannot fit enough signal. You often see loss flatten too early and quality improve slowly across many domains. That matches option A. The other options describe different issues. Exploding norms usually point to insufficient control, not excessive decay. Faster early improvement from high decay is backwards. Communication bottlenecks are a systems issue and not caused by weight decay value itself. So the strongest clue is stable underfitting, not instability.

Excessive weight decay most often shows up as **capacity suppression with stable optimization**. In long pretraining, weights are repeatedly shrunk each step. If decay is too high, representational capacity is damped enough that training loss plateaus early and broad underfitting signals persist. Even when LR schedule, data quality, and systems throughput look healthy. That is option A. B is the opposite failure mode. C incorrectly treats high decay as capacity amplifier. D confuses optimizer math with communication payload mechanics. Interview shorthand: high decay usually fails quietly by limiting fit, not loudly by causing explosions.