Browse Teach Curate

Influence-Consensus Data Selection for Vision-Language Instruction Tuning

A method that ranks training examples in a multitask visual-instruction-tuning mixture by per-task gradient influence scores and then aggregates those rankings by majority voting, selecting a compact subset of examples that are consistently influential across tasks and retain nearly full performance.

Rosie Text

2501.00654

ICONS is a data-selection method for training vision-language models by visual instruction tuning. For each target task it computes first-order, gradient-based influence scores that estimate how much…