EXCISE fixes exclusion queries, boosting 'NOT' accuracy from 29% to 92%

Late-interaction retrieval models like ColBERT struggle with exclusion queries—think "AI tools but not chatbots." Their additive MaxSim scoring promotes documents that match the excluded term, a failure mode the authors call "exclusion inversion." EXCISE solves this entirely on the query side, leaving the frozen index untouched. Two lightweight modules—just 1.5M parameters total—identify the excluded topic and re-embed a 100-document candidate shortlist. A parameter-free rule then demotes any document matching that topic, correcting the inversion without re-ranking the whole corpus.

The results are dramatic. On the ExcluIR benchmark, EXCISE raises exclusion success@10 from 0.058 to 0.691—an 11.9x improvement. Boolean NOT accuracy jumps from 0.25-0.29 to 0.90-0.92, while no-harm nDCG@10 stays on par with the frozen baseline on its strongest backbone. Across six collections and three backbones, EXCISE beats every fine-tuned cross-encoder on 1,860 pooled queries, while those cross-encoders all sacrifice no-harm performance. The team also releases X-BENCH, a tiered benchmark with explicit, implicit, and compound exclusions plus Boolean and no-harm controls, giving the IR community a standardized way to measure this previously under-tested capability.