Unverified50% confidenceBenchmarkExact time
At moderate noise levels and the middle layers of the model, 2 learnable null slots absorbed approximately 90% of the cross-attention mass, while the EOS token's share dropped sharply to approximately 4%.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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66% similarUnverifiedResearch indicates that when the context exceeds a certain length, the model's recall accuracy for content in the middle portion may drop below 50%.61% similarUnverifiedGoogle 内部统计数据表明，E2E 测试与单元测试的数量比例建议维持在 1:10 乃至更低61% similarUnverifiedAfter 15 training epochs on the demonstration dataset, the text classification model achieves an accuracy of 99%, with a total of 110,000 model parameters (16,000 for Attention, 33,000 for FFN, and the remainder for Embedding)61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902316API
curl https://kongchang.com/api/v1/knowledge/claims/902316MCP
get_claim(id=902316)