Unverified50% confidenceFactExact time
DPO 及其变体（如 IPO、KTO）成为 LLM 后训练阶段的主流方法之一
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/7/2026
First Seen
Valid until: 11/5/2026
Sources
Related Claims
VerifiedAlignment algorithms such as DPO (Direct Preference Optimization) reduce training costs while improving the consistency of model performance on open-ended instructions.72% similarUnverifiedHermes模型使用DPO（直接偏好优化）等对齐技术进行训练64% similarUnverified当前主流 LLM 采用指令遵循（Instruction Following）训练方式，会用概率最高的方式补全输出63% similarUnverifiedQwen2采用Yarn外推与双块注意力（DCA）组合方案提升长序列能力，后训练采用离线DPO加在线DPO的组合62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705623API
curl https://kongchang.com/api/v1/knowledge/claims/705623MCP
get_claim(id=705623)