Unverified50% confidenceFactExact time
A model's refusal behavior is concentrated on a specific directional vector in the residual stream, which can be extracted via PCA or the difference-in-means method.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn contaminated benchmarks, models can directly extract answers from parametric memory rather than relying on in-context retrieval, rendering the reported lossless-compression conclusions meaningless.64% similarUnverifiedPCA achieves dimensionality reduction by performing eigenvalue decomposition of the covariance matrix (or SVD on the data matrix), and can only encode second-order statistical relationships (linear correlations) among variables, thus inherently only capable of discovering linear manifold structures.63% similarUnverifiedModel steering influences model output behavior during inference by directly modifying the activation vectors of intermediate layers, without modifying weights.63% similarUnverifiedThe KV Cache mechanism in inference services may introduce intra-batch state dependencies, causing multiple outputs from the same model within a short period to not be entirely independent.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843464API
curl https://kongchang.com/api/v1/knowledge/claims/843464MCP
get_claim(id=843464)