MULTI3IR: A Benchmark for Multi-perspective Multi-domain Multi-modal Information Retrieval
The paper introduces a 104.9K-query benchmark for testing whether retrieval systems cover multiple implicit perspectives, not just one likely answer.
Multi³IR is built from Stack Exchange queries annotated with perspective descriptions across domains and modalities. The authors say current multimodal retrievers show a single-perspective bias on these open-ended queries. They also propose SPIN, a label- and parameter-efficient steering method for embeddings, and report better perspective coverage plus generalization to unseen open-ended IR benchmarks. HF Daily Papers' note
Multi³IR is built from Stack Exchange queries annotated with perspective descriptions across domains and modalities. The authors say current multimodal retrievers show a single-perspective bias on these open-ended queries. They also propose SPIN, a label- and parameter-efficient steering method for embeddings, and report better perspective coverage plus generalization to unseen open-ended IR benchmarks. HF Daily Papers' note
score 5