Researchers have developed DiSPaT, a novel self-play fine-tuning framework designed to enhance anomaly detection in tabular data using large language models (LLMs). This method leverages f-divergence minimization to reduce the distributional gap between real normal data and model-generated samples. DiSPaT's alternating optimization process involves a policy generating synthetic samples and a critic discriminator distinguishing them from real data, thereby improving the model's understanding of normality. Experiments show DiSPaT surpasses existing LLM-based, deep learning, and classical unsupervised methods on various benchmarks. AI
IMPACT This research offers a new approach for improving LLM performance on tabular data tasks, potentially enhancing their utility in real-world anomaly detection scenarios.
RANK_REASON The cluster contains a research paper detailing a new methodology for anomaly detection using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →