🤖 machine learning

Breaking the Reasoning Horizon in Entity Alignment Foundation Models

This paper proposes a novel Entity Alignment foundation model that overcomes the "reasoning horizon gap" in existing graph foundation models by employing a parallel encoding strategy with seed-based local anchors and a merged relation graph to achieve effective, retraining-free alignment on unseen knowledge graphs.

Original authors: Yuanning Cui, Zequn Sun, Wei Hu, Kexuan Xin, Zhangjie Fu

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuanning Cui, Zequn Sun, Wei Hu, Kexuan Xin, Zhangjie Fu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Two Different Libraries

Imagine you have two massive libraries (Knowledge Graphs).

Library A is in New York. It organizes books by color, then by the author's shoe size.
Library B is in Tokyo. It organizes books by the smell of the paper, then by the number of pages.

Your goal is Entity Alignment: You need to find the exact same book in both libraries, even though they are named differently and organized in completely different ways.

The Old Way (The "Memorization" Problem):
Previous computer models acted like a student who memorized every single book in Library A. If you asked them to find a book in Library B, they were stuck. They couldn't do it because they only knew Library A. To help them with a new library, you had to make them study that new library from scratch, which takes a long time and a lot of energy.

The New Idea (The "Foundation Model"):
The researchers wanted to build a "Universal Librarian" (a Foundation Model) that could walk into any two libraries and instantly find matching books without needing to study them first.

The Hidden Trap: The "Reasoning Horizon Gap"

The researchers discovered a major problem when they tried to use existing "Universal Librarians" (Graph Foundation Models) for this task.

These existing models are great at Link Prediction (finding the next book on a shelf). Imagine a path: Book A → Book B → Book C. The model can easily follow this short path.

However, Entity Alignment is different. It's like trying to find a book in Library A that matches a book in Library B. Because the libraries are organized so differently, the path to find the match is incredibly long and winding.

The Analogy: Imagine you are trying to find a specific house in a city by walking from one end of the city to the other, asking every person you meet for directions.
The Gap: The existing models get lost. They try to walk the whole long path (global search) and get confused by the noise. The researchers call this the "Reasoning Horizon Gap." The model's "vision" isn't long enough to see the connection across two different worlds.

The Solution: EAFM (The "Anchor" Strategy)

The authors propose a new model called EAFM. Instead of trying to walk the whole long path, they use a clever shortcut.

1. The "Seed" Anchors (The Meeting Points)
In real-world scenarios, we usually already know a few pairs of matching books between the two libraries. Maybe we know that "The Great Gatsby" in Library A is the same as "Gatsby" in Library B.

The Analogy: Think of these known pairs as Anchors or Meeting Points. They are the only two spots in the two different cities where you know for sure you are standing in the same place.

2. Parallel Encoding (The Twin Walkers)
Instead of starting at the unknown book and walking all the way to the other library, EAFM starts at the Anchors.

Imagine two twins. One starts at the "Gatsby" anchor in Library A, and the other starts at the "Gatsby" anchor in Library B.
They both walk simultaneously (in parallel) toward the unknown books they need to find.
Because they start from the same "meeting point," they don't have to search the whole city. They only need to look at the immediate neighborhood around the anchor. This turns a "long, confusing journey" into a "short, local walk."

3. The Merged Relation Map (The Universal Rulebook)
Libraries have different rules (schemas). To handle this, the model builds a "Merged Relation Graph."

The Analogy: Imagine a master map that doesn't care about the specific books, but only about the types of connections. It learns that "Author" connects to "Book" in Library A, and "Creator" connects to "Volume" in Library B. It realizes these are the same type of rule, even if the words are different. This helps the model understand the structure without getting confused by the names.

4. The Interaction Module (The Final Check)
Once the twins find the potential matches, they don't just guess. They use a special "Interaction Module" to compare the two books side-by-side, looking at tiny details to confirm they are indeed the same.

Why This Matters (The Results)

The researchers tested this new model on many different datasets (different "libraries").

No Retraining Needed: The model was trained on one set of data and then tested on completely new, unseen libraries. It worked immediately without any extra studying.
Beating the Competition: It performed much better than previous models that tried to just "predict links" or models that had to be retrained for every new task.
Robustness: It worked well even when the libraries were huge, messy, or in different languages.

Summary

The paper argues that trying to find matching items across two different knowledge systems by searching the whole system is inefficient and prone to failure (the Reasoning Horizon Gap).

Their solution, EAFM, is like sending two scouts from a known meeting point (the Anchor) to find the target, rather than sending one scout to wander the whole world. By using these local anchors to guide the search, the model can instantly align new, unseen knowledge graphs without needing to be retrained, making it a true "Foundation Model" for this task.

Technical Summary: Breaking the Reasoning Horizon in Entity Alignment Foundation Models

1. Problem Definition and Motivation

Entity Alignment (EA) is a critical task for fusing heterogeneous Knowledge Graphs (KGs) constructed from different data sources. While existing embedding-based EA methods have achieved success, they suffer from a fundamental limitation: lack of transferability. Traditional models operate in a transductive setting, where parameters are tightly coupled to specific entity sets. Consequently, they cannot align unseen KGs without computationally expensive retraining.

Recent advancements in Knowledge Graph Foundation Models (KGFMs) have introduced relative representations to enable transferability across graphs. However, the authors identify a critical barrier to applying KGFMs directly to EA: the "reasoning horizon gap."

The Gap: Unlike intra-graph tasks (e.g., link prediction) where reasoning paths are continuous, EA requires bridging two disjoint graph structures. Directly treating EA as a "sameAs" link prediction task forces models to perform long-range global inference to locate target entities.
The Consequence: This exhaustive global search leads to suboptimal convergence and poor generalization on unseen KGs, as the inference trajectory often exceeds the receptive field of models designed for single-graph reasoning.

2. Methodology: EAFM

The authors propose EAFM (Entity Alignment Foundation Model), a framework designed to bridge the reasoning horizon gap through a parallel encoding strategy driven by local structural proximity rather than global traversal.

2.1 Global Relation Learning via Merged Graph

To address relational-schema heterogeneity, EAFM constructs a merged relation graph ( $G_{rel}$ ) independent of specific entity contexts.

Construction: Nodes in $G_{rel}$ represent all relations from both KGs. Edges are defined based on five topological interactions (Head-Head, Head-Tail, Tail-Head, Tail-Tail, and Inverse) derived from seed EA pairs (anchors).
Structure-Aware RelGNN: A multi-layer Relational GNN encodes $G_{rel}$ to capture high-order semantic dependencies. It employs query-conditioned initialization and attention mechanisms to generate global relation embeddings ( $R_{global}$ ) that serve as a prior for entity encoding.

2.2 Parallel Entity Encoding with Anchor-Conditioned Message Passing

The core innovation of EAFM is the parallel encoding strategy, which transforms long-range global reasoning into local relative matching.

Anchor-Conditioned Initialization: Instead of relying on pre-trained absolute embeddings, the model initializes query entities based on their topological distance to pre-aligned seed pairs (anchors) within a $k$ $k$ -hop neighborhood.
- Entities within the $k$ -hop neighborhood of a query are initialized with a non-zero vector (e.g., all-ones), while others are zero-initialized.
- This process runs simultaneously on both KGs, activating corresponding anchors in both graphs.
Conditional Message Passing: An Entity-aware GNN (EntGNN) propagates information from these local anchors to the query entities. The message passing is translation-based, incorporating the context-aware global relation embeddings derived from the merged graph.
Shortened Trajectory: By initiating reasoning from shared anchors in both graphs simultaneously, the model significantly shortens the inference trajectory, avoiding the inefficiency of exhaustive global search.

2.3 Interaction and Alignment Learning

Fine-grained Interaction: A learnable interaction module constructs a joint vector for source and candidate entities, modeling feature discrepancies and target context (e.g., $v_{st} = [|h_s - h_t| \oplus h_t]$ ).
Bidirectional Classification: The model optimizes a symmetric cross-entropy loss, treating EA as a bidirectional classification task to ensure mutual consistency between aligned pairs.

3. Key Contributions

Pioneering EA Foundation Model: The paper provides the first quantitative analysis of the "reasoning horizon gap" and proposes a foundation model capable of handling arbitrary unseen EA tasks without additional training.
Anchor-Conditioned Message Passing: A novel encoding strategy that leverages seed EA pairs as local anchors to guide information flow, effectively transforming global inference into local relative matching.
Decoupled Interaction Architecture: A scalable parallel encoding architecture featuring a unified relation learner and a learnable interaction module. This design decouples representation learning from specific graph statistics, ensuring robustness against structural heterogeneity and schema distribution shifts.
Superior Generalizability: Extensive experiments demonstrate that EAFM outperforms state-of-the-art baselines and exhibits strong transferability to unseen KGs without retraining.

4. Experimental Results

The authors evaluated EAFM on three benchmark datasets (OpenEA, SRPRS, DBP) covering diverse scales, languages, and structural densities.

Baseline Comparison:
- ULTRA-LP: Applying a standard Graph Foundation Model (ULTRA) directly to EA as a link prediction task resulted in near-random performance (MRR ~0.014), validating the existence of the reasoning horizon gap.
- ULTRA-EA: Retraining ULTRA from scratch for EA yielded moderate results (MRR ~0.395).
- EAFM (Pre-trained): The frozen EAFM model achieved significantly higher performance (MRR ~0.529) on unseen KGs without any fine-tuning, surpassing even the fine-tuned ULTRA-EA baseline.
Robustness Analysis: EAFM maintained consistent accuracy across sparse vs. dense graphs, small (15K) vs. large (100K) scales, and cross-KG vs. cross-lingual settings. Notably, it outperformed fine-tuned baselines in the challenging multi-language setting using topology alone.
Ablation Studies: Removing the parallel encoding strategy caused the most significant performance drop, confirming the necessity of the anchor-conditioned approach. The merged relation graph and interaction module also contributed positively to performance.
Multi-modal Extension: Incorporating entity names via a pre-trained text encoder (BGE) further improved performance, particularly in cross-lingual scenarios, demonstrating the framework's flexibility.

5. Significance

The paper claims that EAFM represents a pivotal step toward developing a universal, service-oriented model for Entity Alignment. By addressing the reasoning horizon gap through anchor-conditioned parallel encoding, the work demonstrates that it is possible to achieve high-performance EA on unseen graphs without the computational cost of retraining. This shifts the paradigm from task-specific, transductive models to transferable, inductive foundation models capable of generalizing across diverse and heterogeneous knowledge graphs.