agentsfactbearishAll six tested LLMs fail to simulate initial human stances themselves and to produce faithful belief updates from self-generated stancesComputation and Language02 Aug 2026http://arxiv.org/abs/2607.28347v1