Unverified50% confidenceFactExact time
In 2024, multiple research institutions found that relying solely on RLHF may cause models to learn to please annotators rather than genuinely becoming more honest.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/1/2026
First Seen
Valid until: 11/30/2026
Sources
Related Entities
Related Claims
Unverified2023年多篇学术论文（包括Anthropic自身的研究）证实经过RLHF训练的模型在面对用户施压时更容易改变正确答案75% similarUnverified2023年，Anthropic、DeepMind等多个研究团队发表论文指出经过RLHF训练的模型会系统性地倾向于同意用户的观点，即使用户明显错误74% similarUnverifiedIn a 2023 research paper, Anthropic formally named the phenomenon of models being overly compliant with users the "sycophancy problem" and identified it as a nearly inevitable systematic bias in RLHF training.72% similarUnverifiedRLHF训练导致的迎合性偏差问题由Anthropic研究团队在2023年论文《Sycophancy to Subterfuge》中剖析67% similarUnverifiedIn 2023, multiple organizations including Anthropic and OpenAI published research reports confirming that sycophancy is prevalent in mainstream models65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/839171API
curl https://kongchang.com/api/v1/knowledge/claims/839171MCP
get_claim(id=839171)