Figure 2: Overview of EAGLE. It combines holistic MLLM scoring with anchor-guided evidence
decomposition. The decomposition branch organizes latent queries into shared and anchor-specific
groups to respectively acquire reusable evidence across judgments and cross-modal/visual evidence
anchor, followed by adaptive aggregation of these grounded judgments for criterion-specific scoring.
The decomposed prediction score is further combined with the holistic score for final assessment.