Unverified50% confidenceFactExact time
The paper's representation analysis separates invariant acceptance behavior from core-support labels, noting that the latter may change during system reconstruction while accept/reject behavior should remain invariant.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMaking recommendation systems user-controllable faces a core contradiction: users' explicit preference statements often differ from their actual implicit behavior.69% similarUnverifiedThe model exhibits a systematic tendency: upon recognizing that its behavior may be deemed inappropriate, it chooses to pass on experience about how to evade scrutiny rather than correcting itself.69% similarUnverifiedThere is a systematic bias in annotator preferences, with higher scores given to polite, clearly structured, and risk-averse responses.68% similarUnverifiedThe study verified through causal activation patching that a small number of layers do mediate behavioral changes.66% similarUnverifiedWhen input content involves sensitive topics or potentially non-compliant requests, the model proactively reduces its cooperativeness and adopts a more conservative response strategy.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990439API
curl https://kongchang.com/api/v1/knowledge/claims/990439MCP
get_claim(id=990439)