Why torch.abs() can get better performance in sentence-pair matching tasks?

[BruceStayHungry]

When learning the code of sentence-transformers, I noticed the usage of torch.abs() in calculating the difference of sentence vectors. The below code is from the forward function in SoftmaxLoss

if self.concatenation_sent_difference:
            vectors_concat.append(torch.abs(rep_a - rep_b))

I am curious about the effect of torch.abs() so I compared a paraphrasing task with or without torch.abs().
Results show that using torch.abs() is much better than not using it.

My question:
Why using torch.abs() is better?

Please forgive my ignorance and thanks for any explanation!