Unverified50% confidenceFactExact time
Both backends score 71% on the same frozen test set, but a case-by-case comparison may reveal that nine cases got worse and nine got better, with the average masking these flips.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
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.70% similarUnverifiedRanked by question difficulty, persistence rate was 91.4% on the 60 questions that all models answered correctly, but only 73.5% on the 52 questions that at least one model got wrong, indicating that sycophancy tracks uncertainty.70% similarUnverifiedAn evaluation suite with 1,800 test cases has long maintained a stable aggregate pass rate of 97%, but the 64-case multilingual slice within it has a pass rate of only 61%69% similarUnverifiedBecause the experiment used temperature=0 and a fixed seed in a single deterministic run, any field flip moves the score by about 2.8 percentage points out of 12 tasks × 3 fields = 36 units, so the gap between 91.7% and 83.3% cannot be directly generalized into a statistically meaningful general advantage.66% similarUnverifiedA single "trusted source" prompt caused 7 out of 8 tested models to reverse between 45% and 88% of their correct answers.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1066524API
curl https://kongchang.com/api/v1/knowledge/claims/1066524MCP
get_claim(id=1066524)