Unverified50% confidenceOpinionExact time
长上下文模型在自我审查场景中因锚定效应可能带来更大偏差，早期生成内容对后续注意力计算影响越显著
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
Codex插件接入Claude Code：双Agent协作审查代码的新范式
bilibili锋芒AI7/12/2026
Related Claims
Unverified提示词中靠前的词汇对生成结果影响更大，因为交叉注意力机制对序列开头的token赋予略高的注意力分数79% similarUnverified不同大模型对长上下文中内容位置的敏感度不同，有些对靠前内容更敏感（首因效应），有些对靠后内容更敏感（近因效应），这影响Gotcha追加位置的生效概率79% similarUnverified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% similarUnverifiedThere is a self-evaluation bias problem when the same model both generates content and reviews it.72% similarUnverifiedThe use of attention weights as a basis for interpretability is controversial, as high attention weights do not always correspond to high causal influence.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/613222API
curl https://kongchang.com/api/v1/knowledge/claims/613222MCP
get_claim(id=613222)