Text-to-SQL models may exhibit recurring error generation patterns that are reproducible under a fixed model and difficult to eliminate through continued training. Existing approaches either perform post-hoc correction at the output level or require access to the target domain’s training data for preference optimization, neither of which directly targets these reproducible local decision tendencies. To address these limitations, we propose Targeted Token-level Unlearning for SQL (TTU-SQL), a machine-unlearning-based framework that uses a probe dataset to detect and correct systematic errors without requiring access to the model’s original training data. The framework comprises three components: (1) AST-based token-level divergence detection that precisely localizes erroneous decision points; (2) a targeted error-type extraction procedure that selects three specific types of forgettable errors based on AST divergence and logit-level behavior; and (3) Part-NPO, which applies negative preference optimization only to localized divergence tokens to suppress erroneous continuations while preserving useful SQL generation knowledge. Across the Spider and BIRD datasets and all evaluated model backbones, TTU-SQL yields consistent execution-accuracy gains over equal-budget training-based repair baselines.

Continual learning for Table Question Answering (TableQA-CL) aims to train models on sequential tasks with limited supervision per task. Existing methods either rely on replaying historical data or assume ideal conditions like gold task boundaries, limiting real-world applicability. We propose TORe (Table Operation Recipe Reconstruction), a rehearsal-free framework that reconstructs historical knowledge through schema-aware reasoning patterns. TORe extracts abstract Operation Recipes from table reasoning processes, identifies fine-grained knowledge deviations between current and past tasks, and employs execution-verified LLM synthesis to generate pseudo-examples reflecting both current and historical reasoning patterns. Experiments on two TableQA-CL benchmarks and multiple backbone models show that TORe consistently improves average accuracy over strong rehearsal-free baselines, remains competitive with replay-based methods, and maintains robust performance under different task orders.

Continual Semantic Parsing (CSP) enables parsers to generate SQL from natural language questions in task streams, using minimal annotated data to handle dynamically evolving databases in real-world scenarios. Previous works often rely on replaying historical data, which poses privacy concerns. Recently, replay-free continual learning methods based on Parameter-Efficient Tuning (PET) have gained widespread attention. However, they often rely on ideal settings and initial task data, sacrificing the model’s generalization ability, which limits their applicability in real-world scenarios. To address this, we propose a novel Adaptive PET eXpert meta-learning (APEX) approach for CSP. First, SQL syntax guides the LLM to assist experts in adaptively warming up, ensuring better model initialization. Then, a dynamically expanding expert pool stores knowledge and explores the relationship between experts and instances. Finally, a selection/fusion inference strategy based on sample historical visibility promotes expert collaboration. Experiments on two CSP benchmarks show that our method achieves superior performance without data replay or ideal settings, effectively handling cold start scenarios and generalizing to unseen tasks, even surpassing performance upper bounds.