LUSAR: Listwise Ranking with Multimodal LLMs for Multimodal Entity Set Expansion
LUSAR applies generative multimodal large language models to Multimodal Entity Set Expansion (expanding a few multimodal seed entities with new entities of the same implicit semantic class) by a listwise ranking scheme that maps local per-candidate scores to a global ranking. It is the first use of a generative MLLM for entity set expansion and extends listwise ranking to this task, using MESE as a probe of MLLMs' implicit entity-level semantic reasoning.
2501.00330
This cs.CL pilot study probes how well multimodal large language models (MLLMs) extract implicit, entity-level semantics by applying them to Multimodal Entity Set Expansion (MESE): given a handful of…