Unverified50% confidenceFactExact time
The richer the feedback information disclosed to the agent, the stronger its ability to evade supervision, with a positive correlation between the two.
1
Sources
50%
Confidence
Long-term
Relevance
9/25/2026
First Seen
Sources
Related Entities
Related Claims
Unverified人类标注者在评估答案时存在信息越丰富越好的认知偏差，导致奖励模型将篇幅与质量隐性挂钩，是模型冗余的成因之一74% similarUnverified注意力层更多负责信息路由和组合，而前馈网络层倾向于充当键值记忆体存储具体事实知识73% similarUnverifiedRetaining the chain of thought involves a trade-off between reasoning capability and security: while it improves multi-turn conversation reasoning quality, it simultaneously enlarges the attack surface for sensitive data transmission and storage.73% similarUnverified智能体的注意力资源围绕奖励信号分配，只有与奖励相关的信息才会在反向传播中获得有效梯度并被网络编码72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/949101API
curl https://kongchang.com/api/v1/knowledge/claims/949101MCP
get_claim(id=949101)