Skip to content

Youtu-GraphRAG: Vertically Unified Agents for Graph Retrieval-Augmented Complex Reasoning

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; Youtu-GraphRAG: vertically unified agents for graph retrieval complex reasoning

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Youtu-GraphRAG proposes a vertically unified agentic paradigm that jointly optimizes graph construction and retrieval via a shared graph schema, rather than treating them as isolated components. A schema-bounded extraction agent, dual-perception community detection, and schema-guided agentic retriever are tightly coupled, yielding up to 16.62% higher accuracy and 33.60% fewer tokens than SOTA baselines across six benchmarks.

Problem

Prior GraphRAG systems optimize either graph construction (e.g., hierarchical clustering via Louvain/RAPTOR) or graph retrieval (e.g., GNN-RAG, HippoRAG) in isolation, creating a semantic gap between the two stages that degrades multi-hop reasoning, especially under domain shift. Additionally, pre-trained LLM knowledge leaking contaminates benchmark evaluation, making it impossible to assess whether answers derive from retrieval or parametric memory.

Method

The framework couples three components through a shared graph schema \(\mathcal{S} \triangleq \langle S_e, S_r, S_{attr} \rangle\) (entity types, relation types, attribute types):

  1. Schema-bounded agentic extraction. A frozen LLM agent extracts triples constrained to \(\mathcal{S}\), reducing open-ended extraction to a structured search space. The schema auto-expands via a confidence-gated update \(\Delta\mathcal{S} = \mathbf{I}[f_{LLM}(d, \mathcal{S}) \odot \mathcal{S}] \geq \mu\) (threshold \(\mu = 0.9\)), enabling domain transfer without manual re-annotation.

  2. Dually-perceived community detection. Entities are encoded by aggregating one-hop triple embeddings (MiniLM-L6-v2). A dual-perception score \(\phi(e_i, C_m) = S_r(e_i, C_m) \oplus_\lambda S_s(e_i, C_m)\) combines Jaccard relation-type overlap (\(S_r\)) and cosine semantic similarity (\(S_s\)) to iteratively merge communities. The result is compressed into a four-level Knowledge Tree \(\mathcal{K} = \bigcup_{\ell=1}^{4} L_\ell\): \(L_4\) = community summaries, \(L_3\) = pivot keywords, \(L_2\) = entity–relation triples, \(L_1\) = attribute facts.

  3. Agentic retriever. A schema-guided query decomposer maps complex queries to atomic sub-queries typed by \(\mathcal{S}\), each targeting node-level, triple-level, or community-level retrieval. The agent operates as \(A^{(t)} = f_{LLM}(q^t, H^{(t-1)})\), alternating forward reasoning and backward reflection with multi-route traversal (entity lookup, triple matching, community filtering, DFS path search).

  4. AnonyRAG dataset. Named entities in documents are anonymized; the "Anonymity Reversion" task requires models to recover original entity names, preventing LLM parametric-knowledge shortcuts during evaluation.

Key Contributions

  • Vertically unified agentic GraphRAG framework binding both extraction and retrieval agents to the same graph schema.
  • Novel dual-perception community detection that jointly optimizes topological connectivity (Jaccard) and subgraph semantics (cosine similarity), building a four-level knowledge tree.
  • Auto-expanding seed schema enabling low-intervention domain transfer.
  • AnonyRAG anonymous benchmark and Anonymity Reversion task to eliminate LLM knowledge-leaking confounds in GraphRAG evaluation.
  • Dual-mode evaluation protocol (open mode vs. reject mode) standardizing fair comparison across GraphRAG methods.

Results

  • Up to 16.62% higher accuracy and 33.60% fewer token costs over SOTA baselines across six benchmarks (HotpotQA, 2Wiki, MuSiQue, GraphRAG-Bench, AnonyRAG-CHS, and one additional dataset).
  • On the construction consumption comparison (Figure 5a), Youtu-GraphRAG achieves the lowest token and time cost among methods with community detection stages (LGraphRAG, E2GraphRAG, RAPTOR).
  • Partial accuracy figures visible: scores in the 67–68% range on several datasets, with Youtu-GraphRAG at 68.41 vs. LGraphRAG 57.41 and RAPTOR 53.0 on one benchmark (full breakdown truncated in provided text).
  • Results hold under both open and reject evaluation modes, indicating robustness of retrieval rather than reliance on parametric knowledge.

Limitations

  • Schema seeding still requires domain-expert-specified initial entity/relation types; auto-expansion cannot start from a fully blank schema.
  • Community detection hyper-parameters (\(\beta=10\), \(\eta=200\), merge threshold \(\epsilon\)) are fixed; sensitivity analysis not visible in the provided text.
  • Evaluation relies on DeepSeek-V3-0324 as a judge for semantic matching, introducing potential judge model bias.
  • The AnonyRAG benchmark covers Chinese-language content (AnonyRAG-CHS), and cross-lingual generalization of the anonymization approach is not evaluated in the visible text.
  • Text is truncated before full ablation results are presented; the relative contribution of each component cannot be fully assessed from the provided excerpt.

Relevance to Agentic AI / LLM Agents

Youtu-GraphRAG exemplifies a tight integration of structured knowledge and agentic reasoning: both the construction and retrieval agents share a common symbolic schema as a grounding artifact, which is a principled solution to the alignment problem between upstream knowledge encoding and downstream query resolution. The iterative perception–reasoning–reflection loop \(A^{(t)} = f_{LLM}(q^t, H^{(t-1)})\) directly instantiates the agent paradigm for multi-hop QA, with the schema acting as an explicit symbolic constraint that steers decomposition and prevents ill-formed sub-queries. The auto-expanding schema mechanism addresses the domain-generalization challenge that is central to deploying LLM agents in open-world settings. For researchers tracking agentic AI, this work demonstrates that binding agents to structured, evolvable ontologies—rather than free-form prompting—substantially improves both efficiency and accuracy in knowledge-intensive reasoning tasks.