LLMs Can Build Better Knowledge Bases for Temporal Reasoning—If Schema Design Comes First

When large language models answer questions about data that changes over time—like historical timelines in Wikipedia infoboxes—their performance hinges far more on schema design than on the query model itself, according to a new study on temporal tabular reasoning.

The Challenge of Temporal Tables#

Answering questions over evolving semi-structured data remains a stubborn problem for current QA systems. Traditional approaches to structured question-answering have trouble when the underlying facts shift across time periods, and extracting queryable databases from messy, timeline-heavy sources like Wikipedia infoboxes adds another layer of complexity.

Schema Design as the Bottleneck#

Rather than directly extracting answers, the researchers reframed the task as building a knowledge base: they prompted large language models to transform Wikipedia infobox timelines into 3NF-compliant relational schemas (a normalized database design), populated those schemas with data, then executed SQL queries against them. The QA accuracy served as a real-world test of how well the schema construction worked.

Their controlled experiment compared three different schema generators paired with six different query models—a 3 × 6 grid. The results were stark:

• Schema source accounted for 79.5% of exact-match variance in QA performance
• Query model choice explained only 1.6%
• Swapping schemas shifted performance by 14.7 to 20.0 points
• Swapping query models under the same schema shifted it by just 4.4 to 12.1 points

Design Principles That Matter#

From these findings, the team distilled three candidate design principles that correlate with better performance:

• Balanced normalization — avoiding over- or under-decomposition of the schema
• Semantic naming — using meaningful, consistent names for tables and fields
• Consistent temporal anchoring — clearly marking when each fact applies

Strong Results#

The best configuration—using Gemini 2.5 Flash to generate schemas and Gemini-2.0-Flash to run queries—achieved 80.39% exact-match accuracy, a (68.89%). An open-weights configuration also reached 79.52%, showing the gains are not locked to proprietary models.