attention-head chain-of-thought data-augmentation decoding hallucination internal-consistency knowledge-distillation large-language-model large-language-models preference-learning reasoning self-consistency self-correct self-correction self-feedback self-improvement self-refine