Overview

• HetaRAG proposes a unified retrieval plane that fuses vectors, knowledge graphs, full‑text indexes, and relational databases to boost recall, precision, and contextual fidelity, with partial code released.
• Think‑on‑Graph 3.0 introduces a multi‑agent MACER mechanism that jointly evolves the query and subgraph during reasoning, reporting stronger performance on deep and broad benchmarks.
• Influence‑guided context selection reframes evidence filtering with a Contextual Influence Value and a surrogate predictor at inference time, showing improvements across eight tasks with code available.
• A five‑level L1–L5 capability framework debuts with aligned benchmarks, finding that multi‑space retrieval plus dynamic orchestration better supports enterprise question answering beyond surface summarization.
• G‑reasoner presents a unified graph‑language approach with a standardized QuadGraph abstraction and a 34M‑parameter graph foundation model, reporting gains over state‑of‑the‑art baselines on six benchmarks.