We reproduce and stress-test the work of Yu et al. (2023), who characterize how language models (LMs) arbitrate between memorized knowledge and contradictory…
机构:IRIT
来源:arXiv 2609.24238 | AI4Papers 论文推荐平台