Tag
DiARC is a method that improves the reasoning ability of large language models on ARC-like tasks by constructing preference pairs from positive and negative samples, outperforming baselines across multiple benchmarks.