In the LLM arms race, the conventional wisdom is the bigger the parameters, the stronger the capability. But the latest research from the StringNLPLAB team at the National University of Singapore is shaking this rule. The team proposes FlexSQL, a Text-to-SQL agent whose core design principle is flexible database interaction: the agent can explore schema structures, check data values, and run validation queries at any time during reasoning, rather than retrieving schema information only once at the start like traditional systems.
FlexSQL generates diverse execution plans to cover multiple query interpretations, and supports both SQL and Python execution modes, flexibly switching based on task type. Its two-layer repair mechanism can roll back from code-level errors to plan-level revisions, while traditional systems can only do post-hoc fixes.
In the Spider2-Snow benchmark, FlexSQL using gpt-oss-120B reached 65.4% score, surpassing GPT-o3 and DeepSeek-R1 which use stronger and larger models. When FlexSQL is integrated as a skill into Claude Code, it achieves over 10% relative improvement.
FlexSQL proves that architectural flexibility may be more important than model scale. For enterprise deployment, this means maintaining high performance while using smaller, cheaper, more efficient models, significantly reducing cost. This work echoes the recent test-time-compute trend: giving the model more reasoning time and interaction freedom often works better than stacking parameters.