DEAL-SQL: Continual Learning for Multi-Dialect Text-to-SQL via Cognition Vectors
This repository is the first authorized GitHub publication of the current manuscript version, released with the explicit permission of the original authors on July 10, 2026.
DEAL-SQL is a teacher-student framework for continual learning across multiple SQL dialects. It separates schema information, general SQL syntax, and dialect-specific syntax, then uses a dialect cognition vector derived from attention-head activations to support adaptation to new dialects while reducing forgetting of previously learned dialect knowledge.
- Anonymous manuscript
- Submission status: NeurIPS 2026 anonymous review
- First authorized GitHub publication: July 10, 2026
- Official repository: https://github.com/AgnesAI-Labs/DEAL-SQL
The original authors have explicitly authorized public distribution of this manuscript through this repository. The "Do not distribute" footer in the PDF is inherited from the anonymous submission template and does not reflect the authors' current authorization for this repository release.
Author names and final citation metadata remain withheld during anonymous review and will be updated after the review process permits disclosure.
Copyright remains with the original authors. Public access does not waive authorship, attribution, or other intellectual-property rights.
When discussing or building on this work, cite the paper title and this repository URL until the final citation is published. Plagiarism, false claims of authorship, removal of attribution, and redistribution that misrepresents this work as another party's original research are prohibited.
See NOTICE.md for the complete publication and attribution notice.