Florian Brand(@xeophon):getting >30 pp improvements by changing to native tool calling, preserving reasoning + setting correct sampling params so many evals have those subtle mistakes which are easy to catch if you know where to look. or just use verifiers + the native harness : )

Florian Brand

@xeophon

evals @PrimeIntellect | open models @interconnectsai

加入 July 2015

793 正在關注 15.9K 粉絲

Florian Brand@xeophon

2026.08.04 07:19

getting >30 pp improvements by changing to native tool calling, preserving reasoning + setting correct sampling params so many evals have those subtle mistakes which are easy to catch if you know where to look. or just use verifiers + the native harness : )