A research team has developed IRIS (Identity Representations from Internal States), a training-free framework that achieves near-perfect accuracy in entity alignment across knowledge graphs using frozen large language models.

The system addresses a core challenge in knowledge graph management: identifying when entities in different databases refer to the same real-world object. Traditional methods struggle with entities described differently across systems, while existing LLM approaches require repeated processing for each new comparison.

IRIS creates what researchers call "iris-like signatures" for each entity by extracting identity-oriented representations from frozen LLMs. These signatures form a shared space where entities can be compared directly through similarity matching, eliminating the need for pair-specific processing or candidate-wise inference.

Benchmark Performance

The framework achieved 100% accuracy on the D-Y-15K V2 benchmark, with scores of 99.38% on DBP-WIKI, 98.31% on ICEWS-WIKI, and 97.99% on ICEWS-YAGO across four established entity alignment datasets.

Unlike existing methods that tie alignment to specific knowledge graph pairs or candidate sets, IRIS encodes each entity once and enables cross-graph alignment through direct similarity comparison. The approach works with frozen LLM backbones without requiring model fine-tuning or training.

The research, published on arXiv by a team including Xinran Liu, Shengtao Li, and Shouqian Shi, demonstrates how frozen language models can provide deeper semantic understanding for entity recognition tasks. The framework distinguishes entities with similar descriptions while recognizing identical entities across heterogeneous data sources.

IRIS represents a shift from auxiliary LLM usage toward creating stable, directly comparable identity representations. The training-free approach could reduce computational overhead for organizations managing multiple knowledge graphs or integrating external data sources.

The researchers tested the framework across two different frozen LLM backbones, suggesting the approach generalizes across model architectures. The work builds on growing research into leveraging pre-trained language models for structured data tasks without additional training requirements.