Joohyung Yun

Agentic retrieval over multimodal documents must commit to a path without exploring every alternative. FiF treats a failed step as feedback: it backtracks to the last decision point and uses the failure to choose a different branch, reaching state-of-the-art accuracy on MultimodalQA, MMCoQA, and WebQA at lower cost.

LILaC indexes a document collection as a two-layer component graph: a coarse layer for fast candidate generation and a fine layer for reasoning across text, tables, and images. A late-interaction traversal retrieves the relevant subgraph, improving both precision and recall on multihop queries.

Early fusion pre-aligns tables with passages and over-retrieves; late fusion aligns on the fly and under-retrieves. HELIOS combines both over a layered graph and adds an LLM reasoning step, giving higher retrieval accuracy on table-text benchmarks with efficient inference.

Top-down schema discovery decides a node's type before seeing its children and gets heterogeneous data wrong. ReCG works bottom-up, clustering leaf patterns and generalizing upward under a minimum-description-length objective, improving schema accuracy by up to 47% while running faster than prior methods.

A three-step framework for the Meta CRAG challenge: answer from parametric knowledge when it suffices, retrieve external evidence only when needed, and verify the generated answer before returning it, cutting hallucination while keeping retrieval cost low.