EXCISE: Query-Side Exclusion for Late-Interaction Retrieval

Late-interaction retrieval models, such as ColBERT, often fail when users include negative constraints (e.g., 'X but not Z'). Because these models calculate relevance by summing the strongest token-level matches, they inadvertently assign high scores to documents containing the excluded topic. This phenomenon, termed 'exclusion inversion,' causes the very documents a user wants to avoid to be ranked at the top of the results list.

The authors demonstrate that this failure is an inherent property of the additive scoring mechanism rather than a lack of training. They show that simply fine-tuning the encoder is inefficient, as it requires re-indexing the entire corpus and often degrades performance on non-exclusionary queries. Instead, the authors propose EXCISE, a lightweight query-side operator that leaves the frozen index untouched. EXCISE uses two small modules (1.5M parameters) to detect the excluded topic and re-embed the top 100 candidates. A parameter-free demotion rule then penalizes candidates that match the identified excluded topic, using an adaptive threshold to ensure that the penalty is only applied when evidence for the excluded topic is statistically significant.

EXCISE consistently outperforms both frozen and fine-tuned baselines across six collections and three different backbones. On the ExcluIR benchmark, it improves success@10 from 0.058 to 0.691. Unlike fine-tuned cross-encoders, which often suffer from 'no-harm' performance degradation on standard queries, EXCISE maintains the performance of the frozen index on non-exclusionary queries. The authors also introduce X-BENCH, a new tiered benchmark designed to test explicit, implicit, and compound exclusion queries with rigorous controls.

This work provides a practical, efficient solution for high-stakes retrieval tasks—such as legal discovery or systematic reviews—where retrieving even a single unwanted document can be costly. By confining the learning to the query path, EXCISE allows developers to add robust exclusion capabilities to existing, large-scale frozen indexes without the massive computational overhead of re-indexing or the risk of harming general retrieval quality.