TAROT: Task-Adaptive Refinement of LLM-prior Graphs for Few-shot Tabular Learning

TAROT is a new GNN-based framework that improves few-shot tabular learning by constructing task-adaptive semantic graphs from LLM-inferred feature relationships. The approach addresses privacy concerns of direct LLM tabular data processing while achieving state-of-the-art performance on few-shot benchmarks through intelligent graph refinement that filters LLM hallucinations.

TAROT represents a meaningful advancement in handling one of machine learning's persistent challenges: learning effectively from limited labeled data in tabular domains. The framework elegantly bridges two complementary approaches by leveraging LLMs for semantic understanding while using Graph Neural Networks for structured reasoning, creating a hybrid architecture that neither fully trusts LLM outputs nor ignores their semantic insights.

The core innovation addresses a critical problem in few-shot learning: feature interactions matter significantly when data is scarce, yet traditional methods miss semantic relationships between variables. By prompting LLMs to infer connections based on task descriptions and feature names, TAROT captures domain knowledge without exposing raw data—a crucial distinction for applications handling sensitive information in healthcare, finance, or regulated industries.

The task-adaptive refinement mechanism is particularly important because it acknowledges LLMs' tendency to hallucinate relationships. Rather than blindly accepting LLM-generated graphs, TAROT actively prunes spurious edges and adds missing task-relevant connections through learning-based refinement, transforming unreliable outputs into useful structural priors.

For practitioners, this approach offers practical benefits: reduced computational overhead compared to methods requiring extensive unlabeled data generation, stronger privacy guarantees than end-to-end LLM processing, and improved predictive accuracy. The framework appears generalizable across tabular domains, from fraud detection to clinical prediction tasks where few-shot scenarios are common. Success here could accelerate adoption of semantic-aware learning methods in enterprise applications where both data scarcity and regulatory constraints are significant barriers.

• →TAROT combines LLMs for semantic inference with GNNs for structured reasoning to improve few-shot tabular learning accuracy.
• →Task-adaptive graph refinement mitigates LLM hallucinations by automatically pruning spurious edges and adding missing task-relevant connections.
• →The framework addresses privacy concerns by avoiding direct raw data exposure to LLMs while capturing semantic feature relationships.
• →Reduced computational overhead compared to traditional few-shot methods requiring additional training on unlabeled or generated data.
• →State-of-the-art performance on benchmarks suggests potential for enterprise deployment in fraud detection, clinical prediction, and regulated data scenarios.