Position Paper on Optimizing LLMs for Self-Consistency

LLM failures like sycophancy and logical errors often stem from evaluating single-output pairs rather than relationship-based reasoning. The authors argue for treating self-consistency as a core optimization framework to improve adversarial robustness and factual coherence.