Contribution & Conclusion

1. A new framework to combine LLM and Graph data.
2. Ten common scenarios to assess the LLMs’ graph understanding ability.
3. Still a long way to go to make LLMs understanding graph data better. The core challenge lays in the intrinsic differences between sequence data and graph data.

Benchmarks (10 scenarios)

结构理解：

size detection;

degree;

neighbor;

attr;

diameter;

cluster coefficient ;

data: citation-net’s 100 sampled small subgraphs;

语义理解：