Unverified50% confidenceFactExact time
This project employs a multi-loss joint training strategy comprising four components: cross-entropy loss, Pull loss, Push loss, and Diversity loss.
1
Sources
50%
Confidence
Long-term
Relevance
9/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic研究员最早在2020年提出预训练损失与计算量之间存在幂律关系（扩展律）63% similarUnverifiedIMGNet提出度量-损失对齐（metric-loss alignment）假设：相似度度量应与训练目标协同设计，而非默认余弦相似度61% similarUnverified训练-部署鸿沟指模型因训练与部署环境的数据分布偏移、软件依赖版本差异或硬件指令集不同而导致的性能下降60% similarUnverifiedUnbalanced expert load is the most typical training challenge in MoE, which can lead to the phenomenon of 'expert collapse' and is typically addressed by introducing auxiliary load-balancing loss or token-dropping strategies.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/837576API
curl https://kongchang.com/api/v1/knowledge/claims/837576MCP
get_claim(id=837576)