File size: 2,221 Bytes
bc1cd35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---
pipeline_tag: graph-ml
---

# GILT: Graph In-context Learning Transformer

GILT is a novel Graph Foundational Model (GFM) that achieves **LLM-free** and **tuning-free** in-context learning on graphs. Unlike existing approaches that rely on Large Language Models or require costly per-graph tuning, GILT reframes few-shot graph learning as a token-based reasoning problem, enabling direct inference on new tasks without any parameter updates.

- **Paper**: [GILT: An LLM-Free, Tuning-Free Graph Foundational Model for In-Context Learning](https://huggingface.co/papers/2510.04567)
- **Repository**: [https://github.com/yiming421/inductnode/](https://github.com/yiming421/inductnode/)

## Key Features

- **LLM-Free**: Works directly with numerical features without text dependency.
- **Tuning-Free**: Adapts to new tasks via in-context learning without gradient updates.
- **Multi-Task**: Unified framework for node, link, and graph classification.
- **Efficient**: Orders of magnitude faster than tuning-based or LLM-based methods.
- **Strong Performance**: State-of-the-art few-shot results across diverse benchmarks.

## Installation

```bash
# Create conda environment
conda env create -f env.yml
conda activate gnn
```

## Quick Start

### Training

To train GILT from scratch with the default configuration:

```bash
python train.py
```

### Few-Shot Evaluation

GILT requires no tuning for new tasks. You can simply provide a few labeled examples at inference time. To evaluate using a pre-trained model:

```bash
# Download checkpoint (replace with local path if downloaded manually)
# Multi-task evaluation
python train.py --use_pretrained_model true --load_checkpoint checkpoints/gilt_model.pt
```

The model performs in-context learning without any parameter updates, directly inferring from few-shot examples.

## Architecture

GILT consists of two main components:
1. **Graph-Native Tokenization**: Converts heterogeneous graphs into unified token representations using PCA-based feature alignment and a deep linear GCN encoder.
2. **In-Context Reasoning**: A Transformer-based module that performs reasoning over contextual tokens using a two-stage attention mechanism and a prototypical prediction head.