Unverified50% confidenceBenchmarkExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
UnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.72% similarVerifiedIn long-chain tasks, if each step has an error rate of approximately 5%, the overall success rate may drop to below 36% after 20 steps.68% similarUnverifiedThe released checkpoint was trained with a chunk size of 16; changing the sampling chunk size to 64 causes HumanEval to drop from 76% to approximately 20%, and GSM8K to drop from a high position to 2.2%.67% similarUnverified实验数据显示，经过5-10轮迭代训练后，模型在低频词汇和长尾知识上的覆盖率下降超过40%，而输出文本在表面流畅性上几乎没有可察觉的退化67% similarUnverifiedThe power-law exponent between test loss and compute is approximately 0.05, meaning that every 10x increase in compute yields approximately a 12% improvement in model performance.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/565267API
curl https://kongchang.com/api/v1/knowledge/claims/565267MCP
get_claim(id=565267)