Hard Negative Mining

Hard negative mining selects or constructs non-matching examples that resemble the correct answer, teaching distinctions that easy negatives may not. Methods include choosing near-boundary items from embedding spaces or model errors, filtering in-batch negatives, and generating synthetic or counterfactual examples with language models or learned samplers. The approach is used in contrastive vision-language learning, search and recommendation, entity matching, and anomaly detection, with evaluations spanning retrieval benchmarks such as BEIR and MS MARCO. The central challenge is calibration: easy negatives add little signal, while overly difficult or mislabeled ones can be false negatives, teach source artifacts, or cause overfitting.