safetyfactbullishLLMs do not sycophantically retreat from established scientific consensus when users signal doubt, showing reactive assertion, surface hedging, or non-response insteadComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01951v1