IGMiRAG: Intuition-Guided RAG Framework

• IGMiRAG is a retrieval-augmented generation framework that integrates hierarchical, intuition-driven memory mining to enable adaptive multi-hop reasoning.
• It employs a Hierarchical Heterogeneous HyperGraph to unify raw documents, atomic entities, composites, and hypernodes into a multi-layer structure for improved fact retrieval.
• The framework achieves state-of-the-art performance on multi-hop QA tasks with lower token consumption and enhanced interpretability over traditional RAG systems.

Intuition-Guided Retrieval-Augmented Generation with Adaptive Mining of In-Depth Memory (IGMiRAG) is a retrieval-augmented generation (RAG) framework designed to address the limitations of classical and graph-based memory structures in LLM systems. IGMiRAG introduces a hierarchical, intuition-driven memory mining approach that mirrors human reasoning, enabling more effective multi-hop and higher-order fact retrieval at variable computational cost, particularly in scenarios demanding deep, compositional inference (Hou et al., 7 Feb 2026).

1. Motivation and Theoretical Basis

Retrieval-augmented generation frameworks supplement LLMs with external knowledge memory, typically retrieved by similarity search over unstructured or chunked document representations. Standard RAG systems concatenate the top-$k$ most similar text fragments, relying on lexical/embedding similarity, which inherently fails to capture multi-hop associations and structured event relationships. As a consequence, such models struggle on questions demanding reasoning across multiple facts or abstraction levels, often producing fragmented or hallucinated outputs. Prior graph and hypergraph-augmented RAG proposals sought to introduce relational and compositional structure, but the memory organization remained misaligned, resulting in computationally expensive, disjointed retrieval (Hou et al., 7 Feb 2026).

IGMiRAG draws from human cognition, particularly intuition-guided reasoning, where individuals rely on rapid, layered exploration of memory structures—determining adaptively how deep and broad to search—prior to detailed analysis. IGMiRAG’s architecture and query parsing procedures incorporate these principles, aiming for unified, granular memory access and dynamic, task-adaptive resource allocation (Hou et al., 7 Feb 2026).

2. Hierarchical Heterogeneous HyperGraph (HHHG) Memory Organization

Central to IGMiRAG is the Hierarchical Heterogeneous HyperGraph (HHHG), which unifies four types of knowledge representations—raw documents ($D$), atomic entities ($N$), pairwise composites ($C$), and multi-entity hypernodes ($H$)—into a single, multi-layer structure: $$\mathcal{G}_{\text{hhhg}} = (V, E),\quad V = N \cup C \cup H,\quad E = CR \cup HR$$

• Vertices:
  • $N = \{ [N_a, De(N_a), A(N_a)] \}$: Atomic entities, with semantic descriptions and attributes.
  • $C = \{ [N_a, N_b, De(N_a,N_b), A(N_a,N_b)] \}$: Pairwise entity composites.
  • $H = \{ [N_a, ..., N_k, De_H, A_H] \}$: Hypernodes for multi-entity concepts/events.
• Hyperedges:
  • $CR \subset N \times C$: Encodes combinations forming pairwise facts ($D$0).
  • $D$1: Encodes high-order events or concepts ($D$2).
• Document Mapping:
  • $D$3: Maps structured memory to raw document sources.

This layered approach enables both bottom-up aggregation (fine-to-coarse) and top-down decomposition (coarse-to-fine), supporting multi-granular and multi-hop inference within a single, navigable structure (Hou et al., 7 Feb 2026).

3. Intuition-Guided Query Parsing and Retrieval Strategy

Prior to explicit memory retrieval, IGMiRAG employs a Retrieval-Strategy Parser (RSP), leveraging an LLM to interpret the user query $D$4 into a six-tuple: $D$5

• $D$6: Rewritten query for clarity.
• $D$7, $D$8: Key entities and keywords for initial matching.
• $D$9: Target memory layer.
• $N$0: Estimated answer fraction at layer $N$1.
• $N$2: Required semantic depth (multi-hop steps).

These signals drive quota allocations: $N$3

$N$4

Where $N$5 is the base quota for seed retrieval, and $N$6, $N$7, $N$8, $N$9 are system hyperparameters. The adaptive adjustment of these quotas reflects human-like intuition in determining search scope and retrieval depth (Hou et al., 7 Feb 2026).

4. Dual-Focus Retrieval and Anchor Identification

The first step in explicit memory access is the identification of high-quality "anchor" memories through two complementary channels:

• BM25 Matching: Keyword matching on entity/name fields ($C$0).
• Dual-Focus Vector Retrieval (DF-Retrieval):
  • Global HNSW Index ($C$1) over all semantic description vectors.
  • Local HNSW Indexes ($C$2) for each memory type.

Candidate quota is sourced from global and local retrieval in proportions determined by $C$3. The union $C$4 undergoes Reciprocal Rank Fusion (RRF), with per-anchor score: $C$5 Relevance scores for document chunks are propagated according to the document relation $C$6, further normalized by vertex degree (Hou et al., 7 Feb 2026).

5. Preference-Aware Bidirectional Diffusion (PABD)

Deep memory mining employs a bidirectional diffusion process simulating deductive human reasoning. Starting with the seed vertices (anchors) $C$7, diffusion progresses for up to $C$8 steps:

• Top-Down Broadcasting: Activation spreads from hypernodes and composites (coarse) to atoms (fine).
• Bottom-Up Screening: Activation propagates from fine-grained nodes to coarse-grained nodes, subject to a dynamic threshold.

At each directed edge step, the update rule incorporates a preference coefficient $C$9 and decay factor $H$0: $H$1 with $H$2, $H$3 weighting forward and backward diffusion. The threshold $H$4 is adjusted per iteration based on activation feedback, trading off exploration and noise suppression. The process halts when no new vertices activate or after $H$5 iterations; activated vertices receive final extended scores $H$6 (Hou et al., 7 Feb 2026).

6. Empirical Evaluation and Comparative Performance

IGMiRAG was evaluated on a suite of QA tasks:

• Simple QA: PopQA
• Multi-hop QA: MuSiQue, 2WikiMultihop, HotpotQA
• Explanatory QA: Mix, Pathology

Metrics included Exact Match (EM), F1 score, and average input token count per query. Table 1 summarizes the main comparison:

Method	EM	F1	Avg Tokens (k)
Baseline LLM	37.6	42.0	–
Naive RAG (Top-5)	53.4	60.6	11.9
NodeRAG (best graph)	53.5	60.9	5.4
Hyper-RAG (best hypergraph)	50.9	57.7	12.9
IGMiRAG	58.3	65.9	6.3

On the MuSiQue and 2WikiMultiHop benchmarks, IGMiRAG achieved state-of-the-art results: $H$7 EM and $H$8 F1 over the closest competitor, with 10–60% lower token consumption. Token cost scaled adaptively to the estimated semantic depth of the question (minimum 3.0k+, average 6.3k+ tokens) (Hou et al., 7 Feb 2026).

7. Impact, Practical Recommendations, and Evaluation

IGMiRAG represents a paradigm shift toward intuition-inspired, structured RAG. By aligning memory structure, query guidance, and adaptive context management, IGMiRAG enables:

• Efficient multi-hop reasoning—higher accuracy on complex QA tasks.
• Adaptive cost scaling—retrieval window and token usage increase only as semantic depth justifies.
• Interpretable strategies—with explicit record of query decomposition $H$9, activated anchors, and diffusion pathway.

For image-grounded IGMiRAG tasks, evaluation should employ claim-centric, modality-agnostic metrics such as MIRAGE INFOF1 and CITEF1 rather than classical ROUGE/BERTScore or text-specific RAG metrics (ALCE, ARGUE, RAGAS), which are insufficient for subclaim-level verification in multimodal settings. MIRAGE decomposes both outputs and reference into unimodal subclaims and verifies each using verifiers (e.g., large vision-LLMs), providing factuality, coverage, and citation support signals at fine granularity (Martin et al., 28 Oct 2025).

Key best practices include using MIRAGE automated variants for rapid diagnostics (via reference precision), combining with classical surface-level measures for efficiency, and, for comprehensive evaluation, employing MIRAGE collection precision/citation scores with VLM-based verification. MIRAGE has demonstrated superior alignment with human judgments compared to alternative sentence-level or embedding-based metrics in multimodal RAG, underscoring its suitability for IGMiRAG evaluation (Martin et al., 28 Oct 2025).

Together, these advances position IGMiRAG as a cost-effective, interpretable, and empirically validated framework for retrieval-augmented generation, enabling practical deployment of LLMs in domains requiring deep compositional and cross-text reasoning.