Why language models abandon true answers under pressure

Original: Decomposing Factual Sycophancy in Language Models: How Size and Instruction Tuning Shape Robustness

Source: arxiv.org ↗