Overview
- A cluster of arXiv preprints published in mid‑July introduces methods and tools that aim to reduce bad or misleading retrieved evidence by preferring source‑aware selection over blind multimodal fusion.
- Researchers released QIMG‑7, a multimodal benchmark that shows polluted image or text retrieval can halve performance and propose SATR, a training‑free gate that compares parametric, text‑only, and full multimodal answers to decide when to fall back.
- Several teams recast retrieval as an active decision process: GRASP trains reinforcement‑learning policies to switch between semantic search, keyword search, and paragraph reading, while EvoGraph‑R1 treats knowledge graphs as self‑evolving structures that agents reshape during multi‑step reasoning.
- Work on efficiency includes SVD‑RAG, which uses SVD on sentence embeddings to build hierarchical trees much faster and with similar retrieval quality, and RAGU, an open GraphRAG engine that ships a compact 7B extractor model to cut cost and run on a single GPU.
- Papers also examine governance risks and practices by integrating FAIR Digital Objects into graph nodes for biomedical QA and by measuring how retrieval can transfer ideological stances into model outputs depending on sampling temperature, signaling a need for auditing and broader validation.