LG AI Research Releases ReSQL Framework That Learns from Failure Cases

In enterprise environments, lightweight language models with 1B–9B parameters are preferred for data security and cost efficiency, but existing models have limitations such as frequent SQL execution errors and a lack of self-correction capabilities.

To address these issues, LG AI Research developed the ReSQL (Retrieval-Augmented Error Reasoning for Text-to-SQL) framework. This framework adopts a closed-loop structure that uses failed cases themselves as training data, rather than relying on prompt-based corrections during the inference stage.

ReSQL's 3-Step Operation

ReSQL automates the process where a single language model generates, diagnoses, and learns from errors.

• Constructing Error Reasoning Data: For failed SQL queries, execution error messages and ground-truth SQL are provided together. Based on this, the model analyzes the behavior of the incorrect query, the root cause of the error, and the correction method in a structured format.
• Generating Training Data: The generated error analysis process is transformed into rich reasoning data that explains 'why it failed,' rather than just providing the correct answer.
• Iterative Self-Improvement: The model is continuously retrained using the collected failure cases and generated error reasoning data. Through this, lightweight models internalize error correction strategies without manually curated supervision, enhancing reliability as enterprise Text-to-SQL applications.

This research was presented at ACL 2026, a major academic conference in the field of natural language processing.