Affixal Negation Training Improves LLM Negation Understanding

Training on the NegCue dataset, which contains 1.8M samples of single-word, multi-word, and affixal negations, shows that affixal negations drive the largest performance gains. This method outperforms standard single-word negation training across five downstream benchmarks.