arXiv cs.AISaturday · May 23, 2026FREE

High Quality Embeddings for Horn Logic Reasoning

A new paper on arXiv (2605.20467) presents methods for creating high-quality embeddings that improve the efficiency of neural-guided logical reasoning. The authors propose three key ideas: (1) generating anchors that are more likely to have repeated terms, which helps the model focus on structural patterns; (2) generating positive and negative examples with a controlled balance of easy, medium, and hard difficulties to prevent overfitting or underfitting; and (3) periodically emphasizing the hardest examples during training to sharpen the model's discriminative ability. The embeddings are trained using triplet loss, requiring anchor, positive, and negative examples. Experiments compare different embedding strategies across various knowledge bases, aiming to identify characteristics that make embeddings well-suited for specific reasoning tasks. The work targets improving the ranking of choices made by logical reasoners, leading to more efficient search for answers. No specific dates, prices, or availability are mentioned as this is a research preprint.