GILT: An LLM-Free, Tuning-Free Graph Foundational Model for In-Context Learning

Researchers introduce GILT, a Graph Foundational Model that enables in-context learning on graph neural networks without requiring large language models or per-task tuning. The approach achieves stronger few-shot performance than existing methods while reducing computational overhead, addressing a critical limitation in deploying GNNs to heterogeneous graph data.

GILT represents a meaningful advancement in making graph neural networks more practical for real-world deployment. Traditional approaches face a fundamental trade-off: LLM-based graph foundational models struggle with numerical features prevalent in large-scale graphs, while structure-based models require expensive per-graph fine-tuning that creates efficiency bottlenecks. This research resolves that tension through a token-based in-context learning framework that operates natively on numerical features without linguistic dependencies.

The broader context reveals an industry-wide challenge. As organizations increasingly rely on relational data—knowledge graphs, biological networks, transaction networks—the inability of existing GFMs to generalize efficiently across heterogeneous graph structures has limited adoption. Current solutions either force data into text representations (inefficient) or demand extensive computational adaptation for each new task (impractical at scale). GILT's tuning-free architecture directly addresses production deployment constraints that have plagued graph-based AI systems.

For developers and organizations deploying graph neural networks, this work meaningfully reduces time-to-inference on new graph tasks, which translates to lower operational costs and faster model iteration. The framework's ability to handle dynamic class semantics from context mirrors successful approaches in large language models but applies them to the numerical and structural domain. This bridges the gap between model capability and practical usability.

The impact on cryptocurrency and blockchain systems could be notable if adopted for transaction graph analysis, fraud detection, or on-chain data interpretation where heterogeneous graph structures are common. However, broader industry adoption depends on validation across diverse real-world datasets beyond the current experimental scope.

• →GILT eliminates the need for large language models or per-task tuning while achieving competitive few-shot performance on graph classification tasks.
• →The token-based in-context learning framework operates natively on numerical features, addressing a critical limitation of LLM-dependent approaches.
• →Tuning-free adaptation enables faster deployment of graph neural networks to new tasks with significantly reduced computational overhead.
• →The unified framework handles node, edge, and graph-level classification tasks through a single mechanism, improving generalization across heterogeneous graph types.
• →Potential applications extend to cryptocurrency transaction analysis and blockchain data interpretation where heterogeneous graph structures are prevalent.