Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· pp. 2698-2707· 0 citations· 17 references
Abstract
Tabular data synthesis must satisfy two complementary quality criteria. At the global level, synthetic records must faithfully reproduce the statistical characteristics of the original dataset, including marginal distributions and inter-variable correlations. At the local level, each individual record must satisfy domain-specific logical constraints such as algebraic identities, conditional dependencies, and mutual exclusivity rules. Existing diffusion-based methods excel at global statistical fidelity but frequently violate hard logical constraints, while LLM-based methods offer stronger constraint adherence but suffer from autoregressive distributional drift that degrades global fidelity. In this paper, we propose LLM-Guided Constraint-Aware Tabular Data Synthesis (LCATS), an end-to-end framework that decouples global distribution modeling from local constraint enforcement. LCATS comprises four modules. The Analysis module uses an LLM to discover candidate constraints, encodes them as executable verification scripts, and validates them against the original data to eliminate hallucinated patterns. The Synthesis module employs a diffusion model to generate statistically faithful synthetic data. The Optimization module repairs constraint-violating records through a constrained weighted optimization problem, which dynamically assigns column-level adjustment penalties to minimize distributional distortion. The Evolution module further mitigates distributional distortion through iterative feedback during constraint-aware optimization. Extensive experiments on seven benchmark datasets demonstrate that LCATS achieves strong overall performance against diffusion-based, LLM-based, and hybrid baselines in ?-precision, ?-recall, and downstream predictive performance, while producing zero constraint violations. The code is available at https://github.com/SJTU-DMTai/LCATS.
The DecoupleMix framework delivers two critical capabilities: guiding what data to collect next and rendering dataset validation a controlled, attributable experiment, and is competitive with strong open-source models trained with substantially larger multimodal budgets.
Jiahao Xie, Zhongbin Guo, Qianle Wang et al.· arXiv.org· 0 citations
Equipping a classifier two-sample test with a gradient-boosted discriminator and decomposing it by controlled permutation into marginal, dependency, and numerical-categorical cross components, each read against a fully factorized reference that destroys all dependency while leaving every marginal intact, and against a real-data oracle.
LAB-Tab is proposed, an LLM-augmented Bayesian network (BN) adaptation framework for source-aware few-shot tabular generation that achieves the best performance at the 10% target-data budget, leads four of the six individual scenarios, and reduces the macro Overall score by 33.8% relative to the strongest baseline.
Zijian Shen, Taijie Chen, Bin Zhou et al.· 0 citations
Symbolic regression (SR) discovers closed-form mathematical expressions from data, offering interpretability beyond black-box models. Existing methods suffer from slow convergence in combinatorial search spaces and lack mechanisms to exploit compositional structure in the data. We introduce SMILE (Sine, Multiplication, Identity, Logarithm, Exponential), a hybrid framework that unifies continuous gradient-based optimization with discrete symbolic recovery through three stages: structural analysis of the data to identify the compositional hierarchy of the target expression, continuous optimization to learn parameters of a network that encodes the target expression using interpretable activations, and symbolic recovery through structured pruning, coefficient optimization, and rounding. This final stage distills the learned network into a compact expression with exact symbolic constants. We evaluate SMILE on SRBench across ground-truth and black-box datasets, with ablation studies validating each component. SMILE achieves the highest symbolic solution rate at the largest noise levels, demonstrating strong robustness where competing methods degrade substantially. It consistently lies on the Pareto front of accuracy versus complexity, recovering significantly simpler expressions in a fraction of the time required by the competing methods.
Mansooreh Montazerin, Antonio Ortega, Ajitesh Srivastava· 0 citations
This paper combines symbolic regression with LLMs for feature engineering (SymboLLM-FE) to solve the dual challenges of poor interpretability and numerous iterations by employing a statistical prior-grounded LLM refinement mechanism and single-digit LLM calls.
Zi-Jian Cheng, Zi-Yi Jia, Zhi Zhou et al.· 0 citations
We study continual pre-training (CPT) as a mechanism for adapting general-purpose large language models to specialized domains: mathematics, instruction, code, and natural text. Using singular value decomposition of weight matrices, we find that CPT leaves singular value spectra largely invariant, with adaptation driven mainly by changes in singular vectors. An analysis of attention-head projection matrices reveals strong, domain-dependent head heterogeneity, which we exploit to define a head importance criterion: up to 60% of head updates can be removed without measurable quality loss. Selectively rewinding low-importance heads to their pre-trained state improves benchmark accuracy by up to 4% versus the fully trained baseline. Finally, we identify domain connectivity - linear interpolation between CPT checkpoints yields smooth domain-quality interpolation without notable degradation on either domain - and release Diffract, an open-source toolkit for scalable spectral analysis of billion-parameter models.
N. Borodin, Maria Krylova, A. Zabolotnyi et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
