Search papers, labs, and topics across Lattice.
1
0
3
1
Decomposing complex, verifiable rewards in LVLM reinforcement fine-tuning provably accelerates convergence and improves generalization, offering a principled alternative to monolithic reward optimization.
