Search papers, labs, and topics across Lattice.
Huawei
3
0
6
ACRL stabilizes RL training while boosting exploration, outperforming traditional methods even under low-precision inference conditions.
MXAttention achieves near-FP16 generation quality in video models while cutting computational costs, revolutionizing efficient attention mechanisms.
RLVR's reasoning gains hinge on high-entropy tokens, revealing a critical inefficiency in uniform reward broadcast that EAPO effectively addresses.
