Abdallah Mohamud Ahmed

تخصص الذكاء الاصطناعي · دفعة 2026

Can We Factorize Text? LLMs as Training-Free Subquadratic Retrievers

Large Language Models (LLMs) are good at judging relevance but have a quadratic attention cost, and a limited context window. CorpusProbe factorizes the document to reduce the context window while also producing a subquadratic computational cost. We factorize a document into K chunks and pass through 2K yes/no single-pass probes, with the first K covering local context and the second K covering global context. Tests on the full corpora of eight BRIGHT long-documents, CorpusProbe achieved the highest NDCG@10 (0.526) and best AUC (0.976) in comparison to six other methods, including naively passing the full document to an LLM (0.967 vs 0.753). The main reason for this disparity was attributed to the LLM being unable to fit the whole document in its context window without crashing. Finally, we tested whether CorpusProbe’s cell scores can accurately localize relevant information within documents via manually annotating 100 tip-of-the-tongue documents from OBLIQ-Bench and found that it outperforms the dedicated ColBERT and frontier-embedding models (50.5 vs 40.0 vs 40.0).

اقرأ التقرير النهائي (PDF)