Search papers, labs, and topics across Lattice.
2
0
5
0
RubricRM adapts evaluation criteria dynamically, leading to significant performance gains in visual generative tasks compared to static reward models.
Even top-performing multimodal judges fail to reliably detect errors in specific modalities, revealing hidden blind spots in their evaluations.
