Unverified50% confidenceFactExact time
LLM在面对开放性问题时倾向于输出训练数据中频率最高的安全表达，即听起来正确但信息密度极低的通用结论
1
Sources
50%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Related Claims
Unverified当输入存在歧义时LLM不会报错，而是沿训练数据中最高频的语义路径继续生成，用统计上最常见的答案填补需求空白75% similarUnverifiedRLHF微调阶段相比预训练在数据量级上通常小2-3个数量级，使安全行为编码为低秩覆写而非底层重塑69% similarUnverified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68% similarUnverified端到端模型的核心优势在于避免模块间信息损失和误差累积，但代价是可解释性低、对训练数据分布敏感、难以应对长程推理任务68% similarUnverifiedDifferential Privacy (DP) clips training gradients and injects calibrated Gaussian noise, allowing the influence of any single training data point's presence or absence on model output to be mathematically bounded.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/114236API
curl https://kongchang.com/api/v1/knowledge/claims/114236MCP
get_claim(id=114236)