reasoningfactbearishAll 12 state-of-the-art LLMs tested suffer significantly from salience bias, with severity scaling with distractor densityComputation and Language02 Aug 2026http://arxiv.org/abs/2607.28478v1