KG-ICL: Graph-Based Universal Reasoning

• KG-ICL is a universal framework that unifies symbolic, neural, and in-context reasoning paradigms over knowledge graphs to execute complex first-order logic queries.
• It employs subgraph selection, query decomposition, and hybrid reasoning approaches to efficiently handle multi-hop logical expressivity and improve transfer across diverse domains.
• Empirical evaluations show that KG-ICL methods outperform baselines in metrics like MRR and Hits, demonstrating robust state-of-the-art performance and cross-domain generalizability.

Graph-based Universal Reasoning (KG-ICL) refers to a suite of frameworks and algorithms that unify symbolic, neural, and in-context (often LLM-based) reasoning paradigms over knowledge graphs (KGs), aiming for broad generalization, multi-hop logical expressivity, and cross-task adaptability. These systems are designed to answer first-order logical (FOL) queries—including complex forms with conjunction, disjunction, existential quantification, and negation—over large, potentially incomplete graphs, sometimes in domains or with relation vocabularies unseen during training.

1. Formalization of KG-ICL: Problem, Logic, and Representational Principles

A knowledge graph is typically formalized as $$\mathcal{G} = (\mathcal{V}, \mathcal{R}, \mathcal{E})$$, where $\mathcal{V}$ is a finite entity set, $\mathcal{R}$ is a finite set of relation types, and $$\mathcal{E} \subseteq \mathcal{V} \times \mathcal{R} \times \mathcal{V}$$ is a set of binary relation edges (Amayuelas et al., 2022, Zhang et al., 22 Dec 2025). Logical queries are generally first-order logic (FOL) formulas, encompassing existential quantification, conjunction, disjunction, and negation, which can be expressed in Disjunctive Normal Form:

$$q[V_{?}] = V_{?} \cdot \exists V_1, ..., V_k : c_1 \vee ... \vee c_n,$$

with clauses $c_i$ built from conjunctions of literals (positive or negative edge predicates).

The universal reasoning objective in KG-ICL is to emit answer entity sets $A(q) \subseteq \mathcal{V}$ for any well-formed FOL query over an input $\mathcal{G}$, requiring models to operate beyond transductive (seen-KG) or fixed-schema single-task settings—for example, supporting inference over new KGs or relations with minimal or no retraining (Cui et al., 2024).

2. Methodological Building Blocks

a. Subgraph Selection and Contextualization:

Extracting a compact, query-relevant context from potentially massive graphs is a core primitive. Most KG-ICL algorithms employ neighborhood retrieval strategies which return a $k$-hop subgraph $S(q)$ containing the minimal set of entities/edges sufficient for evaluating the query. This is often implemented via BFS or top-degree expansion, with heuristics to cap size (beam search, tf-idf ranking over relations) (Zhang et al., 22 Dec 2025, Choudhary et al., 2023). For prompt-based methods, a “prompt graph” centered on a query-related example fact is extracted, including its neighbors and shortest paths of limited length (Cui et al., 2024).

b. Query Decomposition and Logical Operator Handling:

Complex queries with multiple logical operators are decomposed into sequences (chains) of single-operator sub-queries (projections, intersections, etc.), matching the structure of neural computation graphs or LLM stepwise reasoning. For example, $$\exists v. (r_1(e_1,v) \wedge r_2(e_2,v) \wedge r_3(e_3,v))$$ is decomposed into three projections and an intersection. This “chain-of-thought” or “chain decomposition” facilitates modular in-context reasoning and reduces prompt complexity (Zhang et al., 22 Dec 2025, Choudhary et al., 2023).

c. Neural, Symbolic, and Hybrid Reasoning:

Neural approaches embed entities and relations as vectors, parameterizing logical operators (projection, intersection, negation) as small neural networks (e.g., MLPs)—producing a single-point embedding for any FOL query, with answer retrieval via nearest-neighbor search in embedding space (Amayuelas et al., 2022). Symbolic and neurosymbolic methods, such as Tunsr, explicitly construct multi-hop reasoning graphs and perform parallel propositional and FOL forward message passing, enabling both neural propagation and differentiable rule induction (Lin et al., 4 Jul 2025).

d. Prompt-based and LLM-in-Context Models:

A strand of KG-ICL unifies in-context learning and KG reasoning by encoding prompt graphs (local KG fragments) using message passing neural networks and mapping entities/relations into unified token spaces. This approach is parameterized to generalize to unseen KGs by relying on relative topological encodings, masked or abstracted identifiers, and prompt-conditional reasoning (Cui et al., 2024). LLM-centric approaches (e.g., ROG, LARK, R2-KG) use explicit chains of chain-of-thought prompts, natural-language-ized subgraphs, and self-consistency or dual-agent abstention mechanisms for robust inference (Zhang et al., 22 Dec 2025, Choudhary et al., 2023, Jo et al., 18 Feb 2025).

3. KG-ICL Systemic Variants and Architectures

Approach	Contextualization	Reasoning Mechanism	Universalization Strategy
Embedding-based (MLP) (Amayuelas et al., 2022)	k-hop subgraph (latent)	Neural FOL operator MLPs	Operator-agnostic, vectorized logic
LLM-driven (ROG) (Zhang et al., 22 Dec 2025)	k-hop edge list	LLM chain-of-thought in-prompt	Query decomposition and prompt chaining
Prompt-based KG foundation (Cui et al., 2024)	Prompt graph per query	Dual MPNNs (prompt encoding, KG reasoning)	Unified tokenization, zero-shot transfer
Dual-agent (R2-KG) (Jo et al., 18 Feb 2025)	Iterative KG exploration	Low/high-capacity LLMs (operator/supervisor)	Self-consistency, abstention, few-shot
Neurosymbolic (Tunsr) (Lin et al., 4 Jul 2025)	Reasoning subgraph	Dual-channel: propositional + FOL message passing	Hop-wise logic merging, rule induction
LARK (Choudhary et al., 2023)	k-hop subgraph	LLM abstract prompt chain	Entity/relation abstraction, modular prompts

These systems may differ in their base models (neural, symbolic, LLM), context acquisition (retrieval, message passing, prompt graph), logic expressivity (full FOL, negation), and universalization (tokenization, abstraction, plug-and-play prompts).

4. Empirical Results and Evaluation Protocols

KG-ICL frameworks are assessed on benchmarks including FB15k, FB15k-237, NELL995, GRBench, and more than 40 additional KGs spanning transductive, inductive, and fully-inductive regimes (Cui et al., 2024). Metrics include Mean Reciprocal Rank (MRR), Hits@1/3/10, Rouge-L (for QA), GPT4Score (LLM-based correctness), Micro/Samplewise F1, and Coverage (fraction of queries not abstained).

Key findings:

• Neural MLP-based FOL models achieve 5–10% relative gain over BetaE and up to 30–40% over GQE across various datasets, including for queries with negation (Amayuelas et al., 2022).
• LLM-based KG-ICL models (ROG, LARK) set state-of-the-art MRR on 14 FOL query types, with LARK outperforming baselines by 35%–84% average MRR, and gains scaling with LLM parameter count (Choudhary et al., 2023, Zhang et al., 22 Dec 2025).
• Prompt-based KG foundation models generalize across 43 diverse KGs, delivering MRR 0.442 (pretrain) vs. ULTRA-pretrain 0.396 and supervised SOTA 0.351, with robustness to unseen schema (Cui et al., 2024).
• Dual-agent architectures with abstention/post-hoc supervision, such as R2-KG, attain high reliability (HitRate up to 99–100% on answered cases) by abstaining on ambiguous or under-supported queries, at the cost of reduced coverage (Jo et al., 18 Feb 2025).

Ablation studies demonstrate the necessity of prompt graphs, unified tokenization, chain decomposition, and multi-hop context aggregation for best performance (Cui et al., 2024, Choudhary et al., 2023, Zhang et al., 22 Dec 2025).

5. Types of Universalization and Transfer

KG-ICL advances transfer in KG reasoning via several mechanisms:

• Unified Tokenization: By mapping relative graph-theoretic properties (distances from anchors, query-relation identity) to a fixed small token space, models can apply pre-trained architectures to any new KG, even if entities/relations are unseen, as long as small query-centric prompt graphs can be constructed (Cui et al., 2024).
• Plug-and-Play Prompts: Changing only few-shot examples in prompts for LLMs or dual-agent architectures suffices to adapt to new graphs or tasks (e.g., single-label QA, fact verification, or temporal QA) (Jo et al., 18 Feb 2025).
• Chain Decomposition and Abstraction: Decoupling KG retrieval from logical reasoning, with abstracted identifiers in prompts, enables deployment without KG- or task-specific model updates (Choudhary et al., 2023, Zhang et al., 22 Dec 2025).
• Domain-Agnostic Foundations: Empirical results show domain generality across biomedicine, academic graphs, literature, and others—with consistent, robust performance regardless of schema, entity set, or KG density (Cui et al., 2024, Amayuelas et al., 18 Feb 2025).

6. Challenges, Limitations, and Future Directions

KG-ICL faces several technical and practical challenges:

• Graph Incompleteness: Performance degrades when gold-standard answer entities or bridge edges are missing from the retrievable subgraph. This is a primary source of failure in both neural embedding and LLM-driven approaches (Zhang et al., 22 Dec 2025).
• Scalability: Multi-hop, densely connected KG fragments are expensive for both MPNNs and LLM context windows. Sampling, node/edge pruning, and example selection are necessary to cap costs (Cui et al., 2024, Lin et al., 4 Jul 2025).
• Prompt/Operator Instability: For LLM-based methods, prompt length, query abstraction, chain depth, and number of few-shot demonstrations all interact to affect accuracy and cost; special care is needed for complex or compositional queries (Zhang et al., 22 Dec 2025, Choudhary et al., 2023).
• Universal Representation Gap: Fully reconciling rich symbolic logic with continuous embeddings remains an open direction, though dual-channel architectures (as in Tunsr) and prompt-graph tokenizers have narrowed this divide (Lin et al., 4 Jul 2025, Cui et al., 2024).
• Hallucination and Reliability: Even grounded chains can be vulnerable to LLM hallucination, though abstention strategies and dual-agent designs can mitigate this at the expense of coverage (Jo et al., 18 Feb 2025, Amayuelas et al., 18 Feb 2025).

Potential future research includes optimal triple-selection for prompts, dynamic context pruning, joint LLM/KG fine-tuning, richer causal structure exploitation, and learned edge-weighting in KG graph traversals (Kim et al., 2024, Lin et al., 4 Jul 2025).

7. Significance and Impact Across the Literature

KG-ICL frameworks represent a convergence of symbolic, neural, and foundation-model (LLM) paradigms for universal, adaptable reasoning over arbitrary knowledge graphs. By supporting full FOL queries, transferring across domains and schemas, and leveraging both in-context learning and message-passing, KG-ICL sets a new standard for reasoning system extensibility, interpretability, and zero-shot generality (Cui et al., 2024, Amayuelas et al., 18 Feb 2025, Zhang et al., 22 Dec 2025, Jo et al., 18 Feb 2025, Lin et al., 4 Jul 2025). Empirical results confirm state-of-the-art or superior performance across a wide diversity of KG reasoning tasks and settings, establishing it as the dominant approach for future research at the intersection of LLMs and structured knowledge reasoning.