Unverified50% confidenceCautionExact time
GPT-Live在敏感对比问题上表现回避，这是RLHF安全对齐训练的副作用
1
Sources
50%
Confidence
Long-term
Relevance
7/13/2026
First Seen
Sources
Related Claims
Unverified针对GPT-4优化的Prompt未必适用于Claude，源于各模型在预训练数据分布、RLHF偏好及安全过滤策略上的根本差异73% similarUnverifiedA gap may still exist between GPT-Live's demonstration performance and stable performance in real complex environments, and actual user experience awaits large-scale user validation.71% similarVerifiedThe improvement in GPT-5.5's Pass@1 mainly stems from a larger pre-training data scale, more refined RLHF alignment, and improved instruction-following capability.71% similarUnverifiedGPT-5.6 underwent approximately 700,000 A100 GPU hours of red-team testing, with harmful activity blocking approximately 10 times higher than the previous generation.68% similarUnverifiedGPT系列、Claude系列与开源Llama/Mistral系列不仅预训练语料来源不同，其RLHF偏好对齐策略也存在显著差异，因此错误相关性更低67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/498141API
curl https://kongchang.com/api/v1/knowledge/claims/498141MCP
get_claim(id=498141)