GraphARC - Output Graph Edge Count: leaderboard

GraphARC (2026) carries the few-shot transformation format of the Abstraction and Reasoning Corpus to graphs; this question asks for the number of edges of the output graph that the inferred transformation would produce. Accuracy in percent over small test graphs; higher is better.

Metric: Accuracy (%) of answers to questions about the number of edges of the output graph that the inferred transformation produces (the model never sees that output), GraphARC few-shot graph-transformation tasks (21 transformations; the model sees a few input-output graph pairs and a test input graph of 5 to 15 nodes encoded as an adjacency or incidence list), answer compared with the value computed from the true output graph; the OpenAI reasoning models ran one system prompt at medium reasoning effort, the other models four system-prompt variants; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-598#178
2O4 Mini98#368
3O3 Mini94#546
4O1 Mini76#750
5Qwen 3 14B67#1269
6Qwen 3 8B61#1873
7Qwen 3 32B55#972
8Qwen 3 4B54#2348
9GPT-4.1 Nano46#2056
10Qwen 3 1.7B27#5860
11Llama 3.1 8B11#5196
12DeepSeek R1 Distill Llama 8B11#7059
13Llama 3 8B6#5900
14Mistral-7B-v0.24#7596
15OLMo 2 7B4#8320

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=grapharc-output-graph-edge-count · How It Works · Data refreshed daily, snapshot 2026-11-21.