Yingru Li

@richardyrli

RL science, scaling & big runs @xAI @SpaceXSI · Grok 4.2–4.7 | prev. masked importance sampling (MIS) → Trust Region Masking adopted by veRL, slime, miles

following
followers

Cognition introduces SWE-2, claiming scores on par with recent frontier models

Cognition claims up to 70% lower cost than recent frontier models. Modal says Cognition uses its sandbox infrastructure for SWE-2’s training rollouts.

3 Sources, 37d ago, first seen 37d ago

Featured in

Cognition introduces SWE-2, claiming scores on par with recent frontier models

Featured in

Cognition introduces SWE-2, claiming scores on par with recent frontier models