Unverified50% confidenceConceptExact time
The Minimum Description Length (MDL) principle is used to evaluate the compression value of a conjecture — if a new theorem can unify the explanation of multiple known results and thereby shorten the description length of the knowledge base, it is considered to have high discovery value
1
Sources
50%
Confidence
Long-term
Relevance
8/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified知识蒸馏、结构化剪枝和低秩分解等模型轻量化手段可将参数量压缩至原始规模的10%-30%65% similarUnverified投机解码通过小模型预测候选Token、大模型批量验证提升吞吐量；量化技术将权重从FP16压缩至INT8/INT4降低显存占用；MoE架构通过稀疏激活降低单次推理计算量62% similarUnverified规模缩放定律研究表明，增大模型参数量能提升常见知识准确性，但对长尾事实的改善边际效益递减62% similarUnverified只有在LLM推理比embedding加向量检索慢数十倍以上时，语义缓存才能同时实现降本与提速60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/825527API
curl https://kongchang.com/api/v1/knowledge/claims/825527MCP
get_claim(id=825527)