Post

Our paper on self-distillation for training bi-encoders got accepted at #ICTIR2025! By exploiting pretrained encoder capabilities, our approach eliminates expensive teacher models and batch sampling while maintaining the same effectiveness.

The key idea: we can use the similarity predicted by the encoder itself between positive and negative documents to scale a traditional margin loss. This performs implicit hard negative mining and is hyperparameter-free.

Results on BEIR demonstrate that our method matches teacher distillation effectiveness, while using only 13.5% of the data and achieving 3-15x training speedup. This makes effective bi-encoder training more accessible, especially for low-resource settings.