Easy mismatches teach little because a model can reject them without learning subtle distinctions. A hard negative is close enough to a correct result that the model must notice an important difference.
In retrieval training, hard negatives can sharpen ranking, but they must truly be unsuitable for the query. If a supposedly negative item is actually useful, teaching the model to reject it may harm results.
ELI5
A hard negative is a convincing wrong answer used for practice. It is harder to spot than an obviously unrelated answer, which makes it useful for careful comparison.
For example, a search for puppy food might return adult dog food. It is much closer than a bicycle, but it still misses the shopper's specific need.
Why not train only with obviously wrong results?
Obvious mismatches are easy for a model to separate. Similar but wrong examples help it learn finer distinctions.
Can a hard negative be labeled incorrectly?
Yes. An item that seems wrong in one context may be useful in another, so training labels and examples need careful review.
