Adversarially Robust Long-Text Reasoning for Large Language Models with Self-Constructed Negative Samples Xiangchen Song pdf