Certain coverage-based visual token pruning strategies lower Expected Calibration Error compared to unpruned models while keeping accuracy similar on POPE, and pruning reduces ECE on ScienceQA-IMG.
An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Does Visual Token Pruning Improve Calibration? An Empirical Study on Confidence in MLLMs
Certain coverage-based visual token pruning strategies lower Expected Calibration Error compared to unpruned models while keeping accuracy similar on POPE, and pruning reduces ECE on ScienceQA-IMG.