Unverified50% confidenceFactExact time
Across five open-weight checkpoints, after declaring non-target attributes as neutral, the target attribute still leaked in 51% to 81% of items.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the specific experimental case, the overlapping-anchor scheme (n=912) showed the model at 55.2% versus the majority class at 51.1%, a net advantage of +4.1pp, about 2.4 standard errors; the non-overlapping-anchor scheme (n=163) showed 59.5% versus 54.0%, a net advantage of +5.5pp, about 1.4 standard errors.65% similarUnverifiedIn a test involving an unfamiliar game with no manual or rules, the previous generation of AI scored 7.8%, while GPT-6 achieved 99.9% under an agent framework that retains reasoning state.64% similarUnverifiedThe source-anchored variant proposed in the paper further reduces source feature displacement by 47.7% to 83.6% compared to purely learned transfer, while preserving most of the alignment gains64% similarUnverifiedAmong the 100 labels in the classification task, fraud and normal each accounted for 50, the majority-class baseline is 50%, and Gemma 4 31B's 55% is statistically insufficient to prove it surpasses the baseline.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1038825API
curl https://kongchang.com/api/v1/knowledge/claims/1038825MCP
get_claim(id=1038825)