Unverified50% confidenceFactExact time
指令跟随能力依赖于监督微调（SFT）阶段和RLHF阶段的偏好对齐优化
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
Gemini Pro付费用户为何集体失望？五大核心槽点深度解析
Related Claims
UnverifiedSafety alignment is typically achieved through stages such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF); more recent and efficient methods such as direct preference optimization (DPO) have emerged, with alignment goals summarized as the 3H principle: helpful, honest, and harmless.76% similarUnverifiedFunction Calling能力通过SFT阶段构造大量工具调用示例数据获得，结合RLHF优化工具选择准确性73% similarUnverifiedFunction Calling的可靠性高度依赖RLHF对调用时机的精确校准，模型需在训练阶段学习何时调用工具的判断边界71% similarUnverified高质量标注数据是监督学习和RLHF的核心原料，该环节目前仍高度依赖人工完成66% similarUnverifiedRLHF在代码场景中需要人类标注者对代码的可运行性、安全性和符合最佳实践程度进行专项评分66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503705API
curl https://kongchang.com/api/v1/knowledge/claims/503705MCP
get_claim(id=503705)