Rank the three ablations in the ColBERT paper by damage...

The ColBERT paper’s ablation study provides critical insights into the architecture’s performance. Collapsing to [CLS] is the most damaging, proving that fine-grained token-level interaction is essential. Mean-for-max pooling costs 5.2%, highlighting the importance of the aggregator’s design in capturing the ‘aboutness’ of a passage. Finally, dropping query augmentation costs 1.7%, which demonstrates that late interaction makes query expansion nearly free and highly effective. A common mistake is focusing solely on MaxSim as the primary driver, which misrepresents the relative importance of these components.