Unverified50% confidenceOpinionExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
Related Claims
UnverifiedHuman annotators exhibit a cognitive bias toward more information being better when evaluating answers, causing reward models to implicitly link length with quality, which is one of the causes of model verbosity.72% similarUnverified元认知反馈通过奖励置信度与准确率的一致性，在奖励信号中引入认知诚实度这一与内容质量正交的新维度67% similarUnverifiedThe attention mechanism is essentially a dynamic weighted summation, where Query represents what is being looked for, Key represents what can be provided, and Value represents the actual content.67% similarUnverifiedBilibili's recommendation algorithm uses user behavior signals such as likes, coins, favorites, and comments as the core weights for content distribution, with favorites and coins weighted higher than likes.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/718396API
curl https://kongchang.com/api/v1/knowledge/claims/718396MCP
get_claim(id=718396)