publications

2026

1. 

   CANOPY: Adaptive-Granularity Evidence Compression for Multimodal RAG

   Hyojeong Yun^*, Jueun Kim^*, Wook-Shin Han (^*equal contribution)

   arXiv preprint arXiv:2610.00923, 2026

   Abs arXiv Website

   Multimodal RAG retrieves text, tables, images, and videos, but choosing a retrieval granularity does not determine how much context to retain within each item. We introduce CANOPY, a framework for adaptive-granularity post-retrieval evidence compression that represents retrieved items as hierarchies and selects multiple regions at different granularities, requesting targeted follow-up retrieval when the accumulated evidence is judged insufficient.

2. 

   Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering

   Jueun Kim, Sungho Park, Wook-Shin Han

   The Fortieth Conference on Neural Information Processing Systems (NeurIPS), 2026

   Abs arXiv Website

   A central bottleneck in multi-hop QA is that the granularity at which a question is expressed differs from the granularity at which corpus evidence is retrievable. We formulate this as retrievable granularity discovery and introduce Hi-Q, an evidence-conditioned framework for hierarchical query refinement that grows a query tree whose topology is determined by corpus support signals rather than by a fixed decomposition template or a pre-built graph.

3. 

   SPARTA: Scalable and Principled Benchmark of Tree-Structured Multi-hop QA over Text and Tables

   Sungho Park, Jueun Kim, Wook-Shin Han

   The Fourteenth International Conference on Learning Representations(ICLR), 2026

   Abs arXiv Website

   SPARTA is a fully automated SQL-centric pipeline that constructs a tree-structured multi-hop QA benchmark by unifying structured and unstructured evidence from tables and text into a single relational representation.