Unverified50% confidenceFactExact time
High linear separability does not always imply directional influence; that is, a linguistic concept being linearly separable in the representation space does not mean it genuinely affects the model's final predictions.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified线性表征假说认为高级语义概念以线性方向编码在残差流中，是当前可解释性研究的核心假设之一73% similarUnverifiedThe Linear Representation Hypothesis holds that rich linear structure exists in the representation space of LLMs, with specific directions corresponding to grammatical and semantic features and even factual knowledge.72% similarUnverified线性表示假说认为概念以近似线性方式编码在向量空间中，可用向量运算操控语义70% similarUnverifiedThe effects of traditional configuration parameters are usually deterministic, whereas the impact of prompt wording changes on LLM behavior can be nonlinear and difficult to predict.70% similarUnverified提示词中早期出现的词汇因串行依赖结构对后续生成具有不成比例的决定性影响，模型无法回头修正68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/907904API
curl https://kongchang.com/api/v1/knowledge/claims/907904MCP
get_claim(id=907904)