上下文一致性追隨導致錯誤延伸

Risk Description

LLMs tend to maintain internal consistency in generated content. Once an erroneous statement is produced early in generation, subsequent output builds a "self-consistent" narrative around the error, leading to systematic error propagation and deepening. For example, after a first-step reasoning error, the model continues seemingly logical reasoning on false premises, producing "structurally correct but factually wrong" complete arguments.