Justy: Long-document QA keeps hitting the same wall: the context window runs out.
Justy: This paper says the real fix isn’t a bigger prompt. It’s structured reasoning over a persistent database of extracted facts.
Justy: The system is called SLIDERS. It pulls salient information from many documents into relational tables, then reasons with SQL.
Justy: The essential insight is the reconciliation step. [pause] Before querying, it repairs duplicate, inconsistent, and incomplete records.
Justy: It does that using provenance, extraction rationales, and metadata, so local extractions become globally coherent.
Justy: That matters because chunking documents only moves the problem downstream. You still need a clean way to combine evidence at scale.
Justy: The results are strong too, including gains on collections reaching 3.9 million and 36 million tokens.
Justy: Today, read arXiv 2604.22294 and focus on the reconciliation stage, not just the benchmark scores.
Justy: Expect more QA systems to look like databases with reasoning, not bigger text windows.