AI RESEARCH
Anchored Confabulation: Partial Evidence Non-Monotonically Amplifies Confident Hallucination in LLMs
arXiv CS.CL
•
ArXi:2604.25931v1 Announce Type: new We identify a previously unknown calibration property of large language models: providing one confirmed intermediate fact toward a multi-step reasoning chain increases the model's confident-wrong-answer rate before full evidence eliminates it. We call this anchored confabulation: a partial anchor commits the model to confident parametric completion of remaining reasoning steps.