Unverified50% confidenceOpinionExact time
AIME只有30个样本，跑4个种子后以80分对比基线79分并加粗宣称超越，在统计上站不住脚
1
Sources
50%
Confidence
Long-term
Relevance
8/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified即使单步准确率高达99%，在300步生成后完全一致的概率也仅为(0.99)^300≈4.9%64% similarUnverifiedClaude Opus 4.8的编程任务中悄悄放过隐蔽bug的概率仅为上一代的四分之一，假报率从0.40降至0，偷懒糊弄率从25%降至0%64% similarUnverified在Terminal Bench 2.1基准测试上，Ultra模式接近92%，超越竞品的88%表现63% similarUnverified报告测出DiffusionGemma平均每步生成token数达到19.74个，而自回归方法每步只能生成1个63% similarUnverified题量并非越多越好：低于60题的测评难以覆盖多维度、信度不足；但超过300题会因答题疲劳导致后半段数据质量下降。个人自测选100-150题、20分钟内完成的版本较为均衡63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/823652API
curl https://kongchang.com/api/v1/knowledge/claims/823652MCP
get_claim(id=823652)