MINT selects the most-attended image tokens in the second decoder layer, masks the rest, and adds contrastive decoding to reduce object hallucinations in LVLMs.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
MINT: Mitigating Hallucinations in Large Vision-Language Models via Token Reduction
MINT selects the most-attended image tokens in the second decoder layer, masks the rest, and adds contrastive decoding to reduce object hallucinations in LVLMs.