Contexts are Never Long Enough: Structured Reasoning for Scalable Question Answering over Long Document Sets

The core bet SLIDERS makes is that structured intermediate representations are more reliable than asking a model to hold and synthesize sprawling context in memory. Rather than extending context windows further, it treats the problem as a database query problem from the start, which shifts where errors can occur and, importantly, where they can be caught.

This sits in direct conversation with the CLARITY benchmark paper published the same day, which exposed how NL2SQL systems break down on ambiguous or unanswerable queries in multi-turn settings. SLIDERS essentially routes through SQL as a feature, not a fallback, but CLARITY's findings suggest that NL-to-SQL translation is itself a fragile step, especially when user intent is underspecified. If SLIDERS depends on clean, well-formed SQL generation to reason over document-derived schemas, it inherits exactly the failure modes CLARITY documented. The two papers together sketch a tension: SQL gives you auditability and composability over large document sets, but the translation layer between natural language questions and valid queries remains a meaningful weak point that neither paper fully resolves.

Watch whether SLIDERS is evaluated against ambiguous or multi-hop questions that stress the NL-to-SQL translation step specifically. If it holds up on those cases, the structured approach has real legs; if accuracy drops sharply there, the CLARITY failure modes apply directly.