LLM Architecture & TrainingPretraining data causes LLM sycophancy before reinforcement learning, researchers findSalvado•Mar 16, 2026