AI Paper Atlas

Research map / Graph learning / Graph self-supervised learning

Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference

Ben Finkelshtein, Silviu Cucerzan, Sujay Kumar Jauhar, Ryen W White

ICLR 2026 · Graph learning › Graph self-supervised learning · approach: LLMs and graph foundation models · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use Graph-as-Code instead of prompting on long-text homophilic graphso4-mini node classification · Photo / Wiki-CS / Citeseeraccuracy: Photo 85.60 (budget prompt) -> 86.40; Wiki-CS 80.80 -> 82.20Improved
increase neighborhood from 0-hop to 2-hop prompting (ablation)Prompting, short-text homophilic · Coraaccuracy: 64.17 -> 83.43Improved
2-hop prompting on dense or long-text graphsPrompting · Products, Texas, Photo, Wiki-CSaccuracy: TokenLimit (fails)Hurt
give LLM more agency: GraphTool to GraphTool+ to Graph-as-Code (ablation)ReAct-style methods · Coraaccuracy: 74.02 -> 81.40 -> 85.16Improved
use LLM modes on heterophilic graphs (vs label propagation / majority)all LLM modes · Cornell, Washington, Wisconsinaccuracy: Cornell Label propagation 41.74 -> Graph-as-Code 92.70Improved
Graph-as-Code vs label propagation on TexasTexas · Texas (heterophilic)accuracy: 78.90 -> 73.60Hurt
add 2-hop summary prompting (iterative neighborhood summarization) (ablation)Prompting on long-text datasets · Wiki-CSaccuracy: 80.80 (budget) -> 81.70Improved
extend summarization from 2 to 4 hops (ablation)Summary prompting · Wiki-CS / Redditaccuracy: Wiki-CS 81.70 -> 80.90; Reddit 57.60 -> 55.90Hurt
scale LLM from Llama to GPT-5 (ablation)Graph-as-Code · 7-dataset averageaccuracy: 67.80 -> 82.71Improved
enable reasoning mode (ablation)Graph-as-Code (Phi-4) · 7-dataset averageaccuracy: 75.38 -> 83.19Improved
randomly shuffle adjacency matrix (destroy structure) (ablation)2-hop prompt vs Graph-as-Code · Coraaccuracy: 2-hop prompt 83.43 -> 44.50; Graph-as-Code 85.16 -> 56.50Hurt
use Graph-as-Code instead of prompting for shortest pathsynthetic shortest path task · Erdos-Renyi graphs (50 nodes)MSE: 2.82 -> 0.00Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy