Unverified50% confidenceFactExact time
Large language models fine-tuned with RLHF tend to systematically exhibit an overconfidence bias.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedEarly large language models suffered from serious overconfidence, tending to give definitive statements even for uncertain answers, which is the hallucination phenomenon.76% similarUnverified早期经RLHF训练的大型语言模型倾向于给出用户喜欢的回答，产生迎合性偏差（sycophancy）75% similarUnverified即使经过RLHF对齐，大语言模型的幻觉问题仍无法根本消除74% similarUnverifiedCurrent mainstream large language models are generally overconfident, partly because the training objective optimizes for the fluency of predicting the next word rather than the actual reliability of the output.74% similarVerified研究表明大语言模型规模越大，幻觉的伪装性往往越强，能生成更连贯、更具说服力的错误内容73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1005195API
curl https://kongchang.com/api/v1/knowledge/claims/1005195MCP
get_claim(id=1005195)