MemVR is a training-free decoding approach for Multimodal Large Language Models that reinjects visual tokens as key-value memory through the Feed Forward Network when the model exhibits uncertainty during generation.

Authors

• Person: Not available Organization: GitHub
  EdinburghNLP/awesome-hallucination-detection - GitHub

Sources

• EdinburghNLP/awesome-hallucination-detection - GitHub github.com GitHub via serper

Referenced by nodes (1)

• Multimodal Large Language Models concept