Unverified50% confidenceFactExact time
Reflective open-ended questions have a high-entropy output space where the distribution of reasonable answers is dispersed, and intra-group noise masks cross-model signals; whereas low-entropy prompts have a concentrated set of reasonable answers, and models with different training distributions will produce quantifiable deviations.
1
Sources
50%
Confidence
Long-term
Relevance
10/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified从信息论角度看，LLM的输出熵相对较低，倾向于收敛到训练数据的高概率区域73% similarUnverifiedEntropy collapse is especially prominent in long Chain-of-Thought training, because reward signals are only given at the end of the reasoning chain and gradient signals at intermediate steps are extremely sparse.72% similarUnverifiedThe more vague the question, the more the model tends to output statistically common, mediocre answers; the more precise the constraints, the closer the output is to the specific requirement.70% similarUnverifiedWhen a model encounters out-of-distribution inputs, prediction entropy increases, the softmax output tends toward a uniform distribution, and the range of candidate tokens permitted by nucleus sampling expands, giving long-tail anomalous tokens a chance of being sampled.68% similarUnverifiedLLM的幻觉问题源于其训练目标是最小化交叉熵损失而非优化命题真值，是与生成机制深度耦合的固有特性而非偶发缺陷67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/965099API
curl https://kongchang.com/api/v1/knowledge/claims/965099MCP
get_claim(id=965099)