Unverified60% confidenceBenchmarkExact time
Anthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.
2
Sources
60%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
Fable 5对决GPT-5.6 Sol：基准测试、定价与AI权力博弈深度解析
bilibili黑纹白斑马7/3/2026
Related Claims
UnverifiedWhen sampling eight candidates without constraints from the model, 10.2% of inputs cannot produce any valid output under limited training data, and this proportion drops to 0.4% when using the full training data77% similarUnverifiedIn one task continuation experiment, despite the reduced context, the output quality as assessed by human evaluation was actually lower.77% similarUnverifiedMoE架构每次推理仅激活少数专家，训练成本相对同等能力稠密模型大幅降低，10B激活1B相比10B稠密模型训练成本低10倍且推理更快75% similarUnverifiedExperimental results show that rehearsal reduced the observed identity degradation but simultaneously decreased cross-domain transfer performance.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/290466API
curl https://kongchang.com/api/v1/knowledge/claims/290466MCP
get_claim(id=290466)