Beyond Top-K: Replacing Black-Box Retrieval with Interpretable Agentic Operations
The paper argues dense top-k retrieval breaks on numeric, table-heavy documents because context such as units and fiscal years is often outside the retrieved chunk.
The authors test this on a 780-page government financial report where most content lines are table rows. Their READ method avoids embeddings and lets an agent use lexical search, structural navigation, and bounded span reads, producing a replayable search trail. On 51 verified questions, READ answers 58.8% correctly versus 15.7% for dense retrieval, while a top-k agent reaches 27.5%. The paper also says BM25 is statistically indistinguishable from READ, narrowing the claim to embedding-free retrieval rather than agentic search itself. ArXiv · AI/CL/LG's note
The authors test this on a 780-page government financial report where most content lines are table rows. Their READ method avoids embeddings and lets an agent use lexical search, structural navigation, and bounded span reads, producing a replayable search trail. On 51 verified questions, READ answers 58.8% correctly versus 15.7% for dense retrieval, while a top-k agent reaches 27.5%. The paper also says BM25 is statistically indistinguishable from READ, narrowing the claim to embedding-free retrieval rather than agentic search itself. ArXiv · AI/CL/LG's note
score 4