Search papers, labs, and topics across Lattice.
Hunyuan Team, Tencent
4
0
7
Current audio editing models are failing spectacularly, with an Exact Match Rate below 5% in complex tasks, exposing a critical need for improvement.
Current audio-visual models nail unimodal quality but still struggle to make music and dance move together rhythmically, highlighting a key gap TMD-Bench is designed to address.
LLMs can get a 27.8% boost in mathematical reasoning by fusing a hardware-efficient optimal control layer directly into their architecture, enabling planning before prediction.
Forget scaling bottlenecks: DG-PG uses differentiable analytical models to slash gradient variance in cooperative MARL, achieving convergence in a 200-agent cloud scheduling task where standard methods fail.
