Unverified50% confidenceFactExact time
The question set of tw-formosa-bench was filtered against a frozen pool of seven models so that their scores were kept at the passing margin (46.9%–74.6%, median about 54%).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
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.68% 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.65% similarUnverifiedThe accuracy of the seven models tested on PhysMent ranges from 25% to 67%64% similarUnverifiedIn the ScopeBench evaluation, the raw capability of the 8 models ranged from 12.2% to 81.1%, and the scope compliance capability ranged from 34.4% to 86.7%.64% similarUnverifiedExperiments by the Long-Term team showed that using the same model with only adjustments to rule constraints and the memory system raised the score from 52.8% to 66.5%, jumping the ranking from Top 30 to Top 5.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1038584API
curl https://kongchang.com/api/v1/knowledge/claims/1038584MCP
get_claim(id=1038584)