Unverified50% confidenceFactExact time
There is an essential difference between the probability scores (logprobs) output by LLMs and calibrated confidence; large language models are widely known to be overconfident, particularly on out-of-distribution samples.
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe correlation between token-level probabilities (logprob) output by a model and factual accuracy is not strong, and fabricated facts may receive the same high generation probability as true information.75% similarUnverified研究显示，在专业技术领域，LLM 的幻觉率显著高于通用知识领域，因为训练数据中领域专家的纠错反馈相对稀少69% similarUnverifiedMany models still exhibit high confidence in incorrect answers, presenting a confidence calibration problem.67% similarUnverified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66% similarUnverifiedLLM生成阶段对词汇表（通常包含32,000至128,000个token）计算Softmax概率分布，形成总和为1的概率向量66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931407API
curl https://kongchang.com/api/v1/knowledge/claims/931407MCP
get_claim(id=931407)