Task-Specific Learning of Calibration and Discriminative Scoring Weights

Investigate whether learning the calibration weight \(\alpha\) and the Discriminative Token Importance (DTI) scoring weights in a task-specific manner can yield further gains over the training-free RaDiCal framework for visual token pruning in VLM listwise rerankers.

Background

RaDiCal is entirely training-free: its layer-dependent calibration weight α\alpha is derived from normalized attention entropy, while DTI combines query relevance with cross-candidate distinctiveness using fixed scoring operations. The paper does not determine whether these components could be improved by learning their weights for a particular task or retrieval setting. Resolving this question would establish whether task-specific supervision can improve calibration fidelity and ranking quality beyond the fixed, training-free formulation.

References

Second, all components are training-free, and we do not investigate whether learning the calibration weight $\alpha$ or the DTI scoring weights in a task-specific manner could yield further gains.

— From Saliency to Discriminability: Rank-Preserving Visual Token Pruning for VLM Rerankers  (2609.00667 - Liu et al., 1 Sep 2026) in Section “Limitations”