Study reveals visual token pruning in multimodal AI models can improve confidence calibration without sacrificing accuracy

Researchers tested LLaVA-1.5-7B model on POPE and ScienceQA-IMG datasets to measure how token pruning affects model calibration and confidence reliability

SCOPE pruning strategy with pure-coverage settings achieved lower calibration error (ECE) than unpruned models while maintaining similar task accuracy

Reducing saliency weights in pruning consistently improved calibration across all tested token budgets, challenging the assumption that efficiency gains require trading reliability

Multiple pruning strategies evaluated including FastV, random pruning, and saliency-only pruning using metrics like Expected Calibration Error, Brier score, and AURC