Unverified50% confidenceFactExact time
The distributional shift is driven by two key factors: the recipient's prior weight and the candidate evidence skew.
1
Sources
50%
Confidence
Long-term
Relevance
9/8/2026
First Seen
Sources
Related Entities
Related Claims
Unverified位置偏差是偏好数据集中一个被广泛研究的系统性问题，人类标注者评估并排回答时会受展示顺序影响70% similarUnverifiedHuman 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.68% similarUnverifiedexposure bias的根源在于训练与推理的分布不匹配，训练时使用teacher forcing策略而推理时模型依赖自身生成的token67% similarUnverifiedThe study finds that candidate integration occurs in the later layers of the network, manifesting as a structured sequence of steps: accepting the external candidate, increasing its weight, and transferring it to the answer state.66% similarUnverified当选民做出选择时所依据的信息基础发生根本变化，原有的授权链条便可能断裂66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/875701API
curl https://kongchang.com/api/v1/knowledge/claims/875701MCP
get_claim(id=875701)