arXiv cs.CL
·
18 hours ago
EntSQL is a new benchmark for evaluating text-to-SQL systems on enterprise knowledge, containing 1,066 Chinese-English examples across five business domains where SQL generation requires understanding proprietary business documents and internal conventions. The best performing system achieves only 15.9% accuracy on English inputs when long-form documents are provided, demonstrating significant difficulty in grounding SQL generation within enterprise contexts. This benchmark addresses a gap in existing text-to-SQL evaluation by testing performance on realistic enterprise scenarios that depend on private business knowledge rather than just schema and database structure.
arXiv cs.CL
·
18 hours ago
Researchers developed GRID, a grammar-constrained decoding engine that ensures SQL generated by large language models is syntactically valid, respects access control policies, and provides compliance guarantees for enterprise use. The system achieves median per-token masking in 3.6-6.7 microseconds and improves execution accuracy by 13 points on Spider benchmarks at 0.5B model scale. This approach enables reliable SQL generation with enforceable security constraints and auditable decision records.
arXiv cs.CL
·
18 hours ago
Researchers developed PCC-SQL, a system that enforces column-level access control policies in text-to-SQL models by applying per-token logits masking during decoding to prevent policy violations. The system achieved 0% leakage rate and 88.7% coverage on the Spider-CU benchmark while adding less than 10% computational overhead compared to direct prompting. This approach enables safer deployment of text-to-SQL systems across trust boundaries by deterministically preventing unauthorized column access based on how columns are used, not just their presence.
