Unverified50% confidenceBenchmarkExact time
RLCD（Reinforcement Learning from Contrastive Distillation）配方比标准交叉熵落后2.5到3.0个百分点
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic研究团队指出，经过RLHF训练的模型在某些推理基准上的表现会出现1%-3%的性能回退64% similarVerified分类任务中的交叉熵损失本质上是对数似然函数的负值，最小化交叉熵等价于最大化模型对训练数据的对数似然59% similarUnverified困惑度等于测试集交叉熵损失的指数：PPL = exp(loss)，loss=1.77对应PPL≈5.8658% similarUnverified在AG News数据集BERT到DistilBERT蒸馏实验中，可靠性门控关系蒸馏达到94.285±0.054%准确率，RAPID达到94.241±0.025%，均优于交叉熵基线的94.154±0.124%58% similarUnverifiedRLAIF（AI反馈强化学习）用AI自身的批判性评估替代大规模人工标注，相比RLHF将人工标注成本降低了一到两个数量级57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/975540API
curl https://kongchang.com/api/v1/knowledge/claims/975540MCP
get_claim(id=975540)