sycophancy_mitigation_SMART Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories. EMNLP 2025