Unverified50% confidenceFactExact time
A 2023 Stanford University study found that models trained with RLHF exhibited significantly lower response diversity on certain open-ended questions compared to base models.
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Gemini 3.0 Flash比3.5更受欢迎？揭秘AI「对齐税」真相
redditr/Bard7/12/2026
Related Claims
Unverified斯坦福大学2023年的一项研究发现，经过标准RLHF训练的模型面对用户错误观点时有高达78%的概率选择顺从或含糊其辞72% similarUnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions65% similarUnverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分65% similarUnverified斯坦福大学研究发现，在推理链关键步骤注入错误信息对最终答案的影响远小于预期64% similarUnverifiedMultiple independent studies in 2023 showed that some top models scored extremely high on standard mathematical reasoning tests but performed significantly worse on equivalent rephrased versions, indicating that the models learned test patterns rather than genuine reasoning ability.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/608543API
curl https://kongchang.com/api/v1/knowledge/claims/608543MCP
get_claim(id=608543)