GraphWalks' parent-finding task, scored only on prompts from the dataset's longest file, spanning roughly 256K to 1M tokens of context.
unassessed
| Category | long-context |
|---|---|
| Subcategory | multi-hop graph traversal, parent lookup, long context |
| Page status | active |
| Metric | F1 score (node-set overlap) |
| Direction | higher_is_better |
| Unit | F1 x100 |
| Dataset licence | MIT |
| Publisher | OpenAI |
This id covers GraphWalks' "parents" operation, scored specifically on prompts drawn from the openai/graphwalks dataset's longest file, `graphwalks_256k_to_1mil.parquet`, whose prompts range from about 256,000 up to roughly 1 million characters. Given a directed graph written as an edge list of hex-hash node names and a target node, the model must return exactly the set of nodes with an edge leading directly into that target -- its direct predecessors -- excluding the target node itself. Unlike the family's BFS task, which must trace outward across multiple hops, a parents query needs only a single hop, so it isolates whether a model can find and extract the right edges from a long context even when it does not have to chain several lookups together.
A prompt with three worked examples followed by a large directed graph and a "find the parents of node X" instruction, drawn only from the dataset's 256K-to-1M-token file. The model replies with "Final Answer: [node1, node2, ...]" on the prompt's last line.
| Model | Provider | Score | Card as of |
|---|---|---|---|
| Claude Mythos Preview | Anthropic | 97.7 | 2026-04 |