17 upvotes
Paper maps hidden retrieval heads in multimodal LLMs
What is A Broader Look into MLLM's Retrieval Heads?
The study shows which attention heads in multimodal models pull out the right visual or text evidence from long inputs. It expands previous work by examining a broader set of models and tasks.
Who it is for
Researchers and developers building multimodal systems can use the insights to debug or improve evidence retrieval in their models.