https://github.com/sapientinc/HRM
https://arxiv.org/pdf/2506.21734
Nothingburger or really is it a big leap? Seems like it is but haven't read the paper myself, I'm too lazy. Some people have been saying it's one of those situations where they train the models in a way that it performs well in tests for optics but still just in 27 million params?!