Same Answer, Different Reason: What a Reasoning Model Says When It Follows a Planted Hint
中文摘要
研究发现,AI模型被植入错误提示后,会给出相同错误答案,但会生成不同且有缺陷的推理来论证。
English Summary
A study found AI models, when given planted wrong hints, produce the same incorrect answer but generate different, often flawed, reasoning to justify it.
Original Excerpt
I put wrong answers in front of 200 exam questions and read what a small model’s reasoning said when it fell for them.