Sometimes the issue is not reasoning. It is representation and how information moves between components.
In our latest study, we show that changing only the format/representation of a table, while keeping the underlying content identical, can significantly shift model performance.
How RePairTQA Works
To isolate the effect of structure, we hold content constant and vary representation, converting structured tables into paired semi-structured versions.
We introduce RePairTQA and evaluate NL2SQL, simple LLM-based, and hybrid approaches across splits by table size, join complexity, schema quality, and query difficulty.
Key Results
The patterns are instructive for agentic systems. Symbolic NL2SQL methods perform strongly with clean schemas but degrade when structure weakens. LLM-based approaches are more flexible but less precise.
Hybrid systems offer the most balanced behavior across varied conditions. For multi-agent and tool orchestration pipelines, this is directly relevant.
Agents rely on structured tool outputs, JSON schemas, retrieval results, and intermediate state passing. Our findings show that representation choices alone can materially change performance, even when content remains fixed.
A System-Level Decision
In complex workflows, format design is not just a preprocessing step. It is a system-level decision that shapes robustness, coordination, and failure modes.
By Yue Zhang, Seiji Maekawa, and Nikita Bhutani
