HA-DPO reframes hallucination reduction in LVLMs as direct preference optimization over style-consistent positive and negative response pairs, yielding large gains such as 35-point POPE accuracy jumps on MiniGPT-4.
Specif- ically, MiniGPT4-LLaMA2 model is fine-tuned with and without style-consistency control and evaluated on a sub- set of SHR
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2023 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization
HA-DPO reframes hallucination reduction in LVLMs as direct preference optimization over style-consistent positive and negative response pairs, yielding large gains such as 35-point POPE accuracy jumps on MiniGPT-4.