Unverified50% confidenceBenchmarkExact time
In the experiments, no detector reached an 80% true positive rate; the best-performing one, Binoculars, reached only 59% at 200 tokens.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAt a similar flagging rate, a zero-shot decision-model monitor caught only 26% of leaking calls, versus 82% for the linear probe.70% similarUnverifiedIn the controlled GPT-2 experiments, although projection updates were clearly higher in lexical hits, the strict answer score of all experimental groups remained zero.70% similarUnverifiedThe author's criterion: a mutation pass rate below 70% indicates holes in the tests, meaning the probes did not reach far enough.68% similarUnverifiedOn a 50-sample evaluation set, an accuracy change from 74% to 76% corresponds to just one more correct sample, a difference that is statistically almost meaningless.67% similarUnverifiedOnly one intervention slightly improved joint accuracy (the answer being correct and in the target language), by just 1.60 points.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1153418API
curl https://kongchang.com/api/v1/knowledge/claims/1153418MCP
get_claim(id=1153418)