Tag
This paper investigates when multi-agent debate helps or hurts data cleaning, finding that debate degrades generation due to critique-induced confusion but improves error detection. It proposes a debate benefit condition and shows that adversarial separation with code-execution grounding produces the first configuration to significantly exceed single-agent performance on a generative task.
This arXiv preprint challenges the 'Garbage In, Garbage Out' heuristic, arguing that aggressive manual data cleaning can limit predictive performance in high-dimensional tabular data by reducing dimensionality needed to triangulate latent drivers.