GR-Bench

GR-Bench is the graph-structured reasoning benchmark introduced in From Tokens to Topology: When Large Language Models Meet Graph-Structured Reasoning (EMNLP 2026 Findings). It contains 1,800 programmatically generated and oracle-verified instances spanning six task families and three prompt-length regimes.

Evaluation code is available at Ryou-code/GR-Bench.

Dataset structure

Split	Prompt-length regime	Instances
main	below 256K characters	600
regime_256k_512k	256K--512K characters	600
regime_512k_1m	512K--1M characters	600

Each split contains 100 instances from each task family:

• MOTIF: local three-node motif enumeration
• BFS: exact-depth breadth-first search
• PATH: bounded path enumeration
• ART: articulation-point identification
• SUBISO: non-induced subgraph matching
• RULE: synchronous non-monotonic rule execution

The main split is the 600-instance evaluation set used for the paper's controlled cross-model comparison. The other two splits instantiate the same mechanism families at larger graph and prompt scales.

Fields

Each row has four fields:

• id: canonical release identifier of the form grbench_{regime}_{task}_{index}
• task: task-family label
• prompt: complete text-only benchmark prompt
• answer: oracle answer serialized as a JSON string; recover it with json.loads

The Parquet files are lossless conversions of the frozen JSONL benchmark files.

Supplementary metadata

metadata/metadata.parquet provides one row per instance with the canonical id, split, task family, difficulty label, prompt length, task-specific graph statistics, answer size, generation configuration, and source task file. Its 1,800 identifiers align one-to-one with the benchmark rows.

Construction and verification

Instances are generated programmatically from task-specific graph constructions. Gold answers are produced by exact symbolic oracles rather than model outputs or manual annotation. Construction-time checks enforce the task-specific structural constraints, answer consistency, prompt formatting, and split membership described in the paper.

Evaluation

Predictions are parsed as JSON arrays and canonicalized as unordered sets; for nested outputs, each inner list is canonicalized as an unordered tuple. Set-F1 is computed independently for each instance and macro-averaged. Missing or malformed final answers receive zero. The released code implements this protocol and supports these Parquet files directly.

Intended use and limitations

GR-Bench is intended for evaluating graph-structured reasoning under a text-only interface. It is synthetic, English-only, and focused on the six mechanisms listed above; results should not be interpreted as a comprehensive measure of general reasoning or performance on naturally occurring graphs. The dataset contains no human-subject or personally identifying information.

Licenses

The dataset is released under the Creative Commons Attribution 4.0 International license. The evaluation code is released separately under the MIT License.

Citation

@inproceedings{zhang-zeng-2026-tokens,
  title     = {From Tokens to Topology: When Large Language Models Meet Graph-Structured Reasoning},
  author    = {Zhang, Qihang and Zeng, Liang},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}