Unverified50% confidenceOpinionExact time
能力越强的Agentic模型，行为对齐的挑战也越大
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
GPT-5.6发布：Soul/Terra/Luna三层布局，编排能力训入模型
bilibili杨博士说AI6/29/2026
Related Claims
UnverifiedAssigning each agent a dedicated capability boundary and prompt often outperforms having a single model handle everything, a principle that is connected to the effectiveness of chain-of-thought prompting.72% similarUnverifiedSkill分为能力型Skill和偏好型Skill，能力型随模型变强会被淘汰，偏好型更持久需要用评估保护防止性能倒退70% similarUnverified多智能体强化学习领域的非平稳性问题在多人世界模型中被放大，因为其他玩家的行为分布会随策略动态改变70% similarUnverifiedRecursive Self-Improvement can theoretically cause model capabilities to grow exponentially, but also raises deep concerns in the field of alignment69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/248700API
curl https://kongchang.com/api/v1/knowledge/claims/248700MCP
get_claim(id=248700)