Text-to-SQL systems answer natural language questions over structured data, but can fail before SQL generation by linking question mentions to the wrong table, column, value, or relation. Even valid SQL can then return an incorrect answer. This raises a central question: What pre-SQL information helps a SQL generator, and when does additional information cease to help? We investigate this question on TempTabQA-C and BIRD. Among the audited SQL Static failures, grounding errors form the largest category on both benchmarks. We introduce GRIPS (Grounding Representation and Input Processing for SQL generation), a three-stage method that derives structured pre-SQL information to make these decisions explicit before query generation. Across GPT-5-nano and Llama-70B, GRIPS-Adaptive improves over SQL Static in 18 of 20 settings, with an average accuracy gain of 11.8 percentage points, while reducing model-call token usage by 47.8%. Our controlled analysis shows which grounding information can improve SQL generation, whereas adding more pre-SQL information does not consistently help. By separating question-input grounding from SQL generation, GRIPS makes text-to-SQL more accurate, efficient, and easier to diagnose.
Figures & tables
Example 1: TempTabQA-C, Hard
C:
Darrell Pace
Personal Information: Birth Date = 1956-10-23
Olympic Games: Individual MedalGold ∣ 1976 ∣ Montreal; MedalGold ∣ 1984 ∣ Los Angeles
Pan American Games: Team MedalGold ∣ 1979 ∣ San Juan; MedalGold ∣ 1983 ∣ Caracas
Pan American Games: Individual MedalGold ∣ 1983 ∣ Caracas; MedalSilver ∣ 1979 ∣ San Juan
Q:
How many medals did Darrell Pace win in his/her twenties?
Table 1: Examples from TempTabQA-C and BIRD where C: Table Context; Q: Question; A: Answer; E: Evidence.
Figure 1: SQL Static failure breakdown by diagnostic class and subcategory. TempTabQA-C grounding errors concentrate on schema-slice and cardinality (G5/G3); BIRD on filter and join grounding (G2/G4).
Figure 2: Overview of GRIPS pipeline. Stage I maps the question and context to a structured pre-SQL record using a frozen benchmark-specific rule-set. Stage II combines the original input, the Stage I record, and SQL demonstrations to generate SQL query. Stage III normalizes and executes the query to produce the final answer.
Table 3: Stage I outputs for the running examples in Table 1 .
Figure 3: Controlled comparison of Stage I information under a fixed Stage II generator. The first four conditions vary the amount of information; the last three vary its source. G+P denotes guided representation with prediction_trace , and G+FK denotes guided representation with verified foreign-key paths.
Figure 4: Share of SQL Static failures rescued by GRIPS , by diagnostic class and subcategory.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Failure composition by diagnostic class after re-auditing SQL Static errors on TempTabQA-C (Easy/Medium/Hard) and BIRD (Simple/Moderate/Challenging).
Figure 6: Concrete Stage I–Stage II projections for the examples in Table 1
Figure 7: Stage II prompt structure. The benchmark input and Stage I information vary by condition; the task, demonstrations, current question, and output contract are held fixed in controlled comparisons.
Figure 8: Share of SQL Static failures rescued by GRIPS , by diagnostic class and subcategory.
Control
Condition
Definition
Illustrative example
Shared
Without Stage I information
The auxiliary Stage I block is empty. The question, schema/evidence, demonstrations, and SQL-generation instructions remain unchanged.
Empty payload: {} .
Shared
With guided representation
Supplies recognized question–input matches with their semantic types and normalized values. No profile summaries are included.
Column: element Tables: bond , atom .
Content
With representation + profile
Adds deterministic counts, structural scope, priority, target determinacy, and coarse difficulty summaries. This is the deployed GRIPS handoff.
Representation: element , bond , atom Implicated tables: 2 FK span: 2 Target: clear; difficulty: moderate.
Content
With expanded profile
Adds internal pattern, count/repetition, requirement, and compressed-signature fields to the deployed handoff.
Pattern: column over two tables Repeated table mentions: 4 Requirement: 5-tuple Signature: repeated-table doublet.
Source
With adapted DIN links
Supplies schema links produced by the DIN-SQL linking stage to the common Stage II generator. This is an adapted intermediate output, not native end-to-end DIN-SQL.
Supplies the MAC-SQL selector’s table and column decisions to the common Stage II generator. The base schema remains available.
Keep: atom , bond Select from connected : bond and atom IDs Drop: molecule .
Appendix
Table 17: Reader-facing definitions and examples of the controlled Stage I information conditions. Examples are condensed from the BIRD question “What elements are in the TR004_8_9 bond atoms?” The first two conditions are shared across comparisons. Content conditions cumulatively add information, whereas source conditions substitute alternative Stage I information under the same Stage II generator.
Prompting-based (i.e., non-fine-tuning) Text-to-SQL methods, where underlying large language model parameters are not changed for the task, face three problems: (i) relying on coarse-grained schema information that may not reveal the fine-grained relationships needed to distinguish ambiguous columns, (ii) failing to capture recurring SQL-generation failures, and (iii) suffering from omission or hallucination of components in complex questions. This paper develops DexterSQL, a prompting/non-fine-tuning-based Text-to-SQL system that improves SQL generation with three novel components: (i) deep schema explorator that identifies ambiguous columns, analyzes their individual and joint data distributions to uncover their relationships and the distinct role of each, (ii) database-agnostic rule creator that mines mismatches between generated and gold SQL only on the training database and converts them into database-agnostic corrective rules that capture recurring LLM failure patterns; and (iii) multi-path SQL generation that introduces a dependency-tree-based intermediate representation that uses the question's sentence structure to guide its decomposition into an SQL skeleton for final SQL generation. DexterSQL achieves a higher accuracy compared to the state-of-the-art using both open-source/weight and closed-source/weight models. Particularly, DexterSQL shows a high improvement of at least 5.5% using an open-weight model (GPT-OSS-120B) on BIRDDev, with total accuracy 70.4%. DexterSQL also shows better improvement of at least 1.4% using closed-weight models, with total accuracy 72.1% and 72.9% on BIRD-Dev with GPT-4o and GPT-5.2.
Anik Pramanik, Murat Kantarcioglu, Vincent Oria +1
New Jersey Institute of Technology, USA. · Virginia Tech, USA.
Text-to-SQL enables natural language access to databases, and recent LLMs have substantially advanced its capabilities. Existing benchmarks such as Spider, BIRD, and Spider~2.0 evaluate schema generalization, large-scale databases, and realistic workflows, but largely overlook enterprise scenarios where SQL generation depends on private business knowledge, such as internal metrics, reporting conventions, and organizational rules. We introduce EntSQL, an enterprise-oriented Text-to-SQL benchmark for evaluating long-context grounding over proprietary business documents. EntSQL contains 1,066 aligned Chinese-English semantic examples across five business domains, with most examples requiring domain knowledge beyond the question and schema and involving complex SQL structures. On English inputs, the best evaluated system reaches only 15.9% when long-form documents are provided, highlighting the difficulty of grounding SQL generation in enterprise knowledge.
Chengxi Liao, Tao Xu, Zulong Chen +7
1HKUST (GZ) · 2Alibaba Group · 4Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) +1
Text-to-SQL maps natural language questions to executable SQL queries. Modern databases often contain large and complex schemas, making schema linking a critical step for accurate SQL generation. Existing methods either rely on full-schema generation, which leaves schema linking implicit within a large search space, or use a separate retriever trained with static gold-column supervision, whose targets may be suboptimal for the current generator policy. To address this issue, we propose Adaptive Co-optimization via Empirical Credit Assignment for Text-to-SQL (ACE-SQL), a reinforcement learning (RL) framework that jointly optimizes schema retrieval and SQL generation under execution feedback. ACE-SQL constructs an online column-set pool from generator rollouts and derives adaptive on-policy retrieval targets from the column set most frequently associated with execution-correct rollouts. This induces bidirectional adaptation, where the retriever adapts toward column sets that the generator can execute correctly, while the generator adapts to the retriever's evolving schema selections under execution feedback. With approximately 3k synthetic Text-to-SQL question-database pairs for RL training, ACE-SQL achieves 65.3% greedy execution accuracy on BIRD Dev while using 0.93k output tokens per query. The repository is available at https://github.com/xbchen1/ACE-SQL.
Xiaobing Chen, Ai Jian, Eryu Guo +1
Harbin Engineering University, Harbin, China · Harbin Institute of Technology, Harbin, China · Beijing University of Posts and Telecommunications, Beijing, China