Expired50% confidenceBenchmarkExact time
The worst-ranking finalist had a KL divergence 2.3 times that of the optimal candidate and suffered a 7% loss in mathematical capability.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/20/2026
First Seen
Valid until: 10/18/2026(expired)
Sources
Related Claims
Unverified消融实验显示，若继续训练主干通用知识掉10%、数学掉17.8%；若两塔共享权重则通用任务掉26%，其余指标全线崩溃65% similarUnverifiedThe optimal variant showed only a 1.83% decline on the GSM8K mathematical reasoning benchmark, while the MMLU general knowledge benchmark remained unchanged.65% similarUnverifiedINT8 quantization loses approximately 0.5% accuracy while halving VRAM usage; INT4 quantization compresses VRAM to 1/4 of the original with an accuracy loss of approximately 1–3%.62% similarUnverified报告同时给出中位数KLD和99.9%分位KLD两个数值，前者反映典型情况，后者捕捉极端长尾错误61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/568116API
curl https://kongchang.com/api/v1/knowledge/claims/568116MCP
get_claim(id=568116)