Unverified50% confidenceBenchmarkExact time
Compared with questions the untrained agent posed to the same expert, post-training questions disclosed 3.7 times fewer private entities and 1.4 times less verbatim code.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/20/2026
First Seen
Valid until: 1/18/2027
Sources
Related Entities
Related Claims
UnverifiedThe 44% reduction in inference tokens is a self-reported figure from the author and has not been verified by third-party reproduction; the original materials do not disclose the data construction method, training algorithm, sampling scale, or benchmark test details.71% 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70% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.70% similarUnverifiedUnder held-out interface perturbations, the accuracy of the post-trained Agent drops by nearly 25%.70% similarUnverifiedIn an experiment with a 32-expert miniature character model, when Token Choice routing had no load balancing loss, a single expert swallowed 57% of the tokens in the early stage of training, and 17 of the 32 experts were almost ignored.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074866API
curl https://kongchang.com/api/v1/knowledge/claims/1074866MCP
get_claim(id=1074866)