Search papers, labs, and topics across Lattice.
2
0
5
0
Ditch the slow "think-first-then-translate" paradigm: ReflectMT internalizes reflection, delivering faster and better machine translation in a single pass.
Supervised fine-tuning can be dramatically improved by explicitly encouraging exploration of low-confidence data and suppressing high-confidence errors, leading to sustained gains in mathematical reasoning even after extensive RLVR training.
