Unverified50% confidenceFactExact time
Linear concept erasure methods (such as LEACE) can guarantee that no linear classifier can predict the sensitive attribute after erasure, but they cannot prevent MLP probes from recovering the erased information.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified线性探测器可能无法捕捉非线性编码的信息，因此内部探测的'未检测到'可能是探测方法本身的局限而非表征真的缺失66% similarUnverifiedIt should be assumed that system prompts will inevitably leak; no confidential information should be placed in system prompts, and output filtering should be added for canary tokens and sensitive strings.64% similarUnverifiedIf linear calibration is allowed, the correspondence can be recovered almost perfectly, indicating that the information in the latent space exists, it is just not organized into a directly interchangeable form.62% similarUnverifiedEven when the scorer is hidden, frontier models can still reason about and speculate on it, making simply hiding the scoring logic an unreliable way to prevent gaming.62% similarUnverifiedLinear readability means that an attribute is distributed roughly along a fixed direction in the representation space, so downstream computation in the model needs only a single weighted sum to extract it, at low cost and with high stability; if the attribute can only be recovered through multi-layer nonlinear transformations, this indicates that it is more deeply entangled in the representation and harder to use effectively during fine-tuning.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098794API
curl https://kongchang.com/api/v1/knowledge/claims/1098794MCP
get_claim(id=1098794)