Navigate the graph.
Find the answer.
Read a multi-hop question, start at its topic entity, and choose each relation and neighbor yourself. The number of hops, reference path, and answers remain hidden until you finish.
Select a dataset
Choose a dataset, then start a challenge.
You are the navigator
This challenge tests human navigation through a knowledge graph, not pretrained MINERVA. At each step, pick a relation and one of the entity neighbors reachable along that edge. Submit when you think the current node answers the question.
Select Canonical question for the original wording or Random paraphrase for a randomly chosen alternative from the released Question-Paraphrased column. The wording stays fixed for the entire attempt, including retries.
There is no required number of steps displayed. A 12-step safety limit prevents endless wandering.
Real edges. Hidden references.
Kinship and MQuAKE-ST load their graph from Hugging Face. The encoded MetaQA dataset loads from the public GCS ZIP archive. Metadata labels are shown when available. MetaQA navigation allows inverse edges and provides answer checking without reference-path scoring.
Note: MQuAKE-ST and MetaQA can require sizable downloads and browser memory. The graph and MetaQA archive are reused while the page stays open. Kinship and MQuAKE-ST also cache sampled QA rows to reduce repeated Dataset Viewer requests. MetaQA has no released reference paths or paraphrases.