Search papers, labs, and topics across Lattice.
2
0
4
0
Monolithic video evaluation dilutes critical action cues, but chunking interactive rollouts into action-aligned visual evidence allows targeted reward models to outperform GPT-5.5 at scoring world model dynamics.
CapRL++ redefines caption quality through utility, enabling models to produce high-fidelity descriptions without the constraints of traditional supervised fine-tuning.
