Unverified50% confidenceBenchmarkExact time
从第160轮到第290轮，MSE损失从0.0690降至0.0652，130个epoch内仅下降约0.006
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedHyperSAE将MSE重建误差从4.57降至4.12（下降9.8%），交叉熵恢复率从75.5%提升至78.9%66% similarUnverified整个训练过程中，flow-matching 损失仅从 0.805 缓慢下降到 0.75463% similarUnverified课程训练过程中损失值从大约3.36下降到0.32左右63% similarUnverified第290轮时损失出现反弹（+0.002），表明模型可能已接近当前配置的性能上限61% similarUnverified代码执行方案将相同任务的token消耗从15万降至2000，减少了98.7%60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/910713API
curl https://kongchang.com/api/v1/knowledge/claims/910713MCP
get_claim(id=910713)