GraphInstruct (Generation): leaderboard

GraphInstruct (2026, a graph-generation benchmark unrelated to the 2024 GraphInstruct graph-reasoning set) tests whether language models can generate graphs that satisfy natural-language instructions, from format and single constraints to joint constraints, numerical attributes, domain semantics and multi-step edits, over 800 instructions. The quality score, 0 to 100 and higher is better, is computed by code from validity, distribution match, reference similarity and constraint satisfaction; this board uses zero-shot prompting.

Metric: Quality S_total (0-100; the paper's 0-1 score times 100): level scores weighted 0.05, 0.10, 0.15, 0.20, 0.25, 0.25 for L0 to L5, each level score a weighted mix of structural validity and distribution match, reference similarity and instruction-constraint satisfaction checked by code (no LLM judge; efficiency excluded), over the 800 instructions of GraphInstruct (instruction-following graph generation in six constraint levels); zero-shot prompting; five generations per instruction at temperature 0.7 with the InstructGraph code-style prefix; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.685.9#173
2Claude Sonnet 4 (20250514)83.4#405
3Qwen 3.5 122B A10B83.1#346
4Qwen 3.5 397B A17B82.9#264
5GPT-4o82.7#751
6GraphInstruct DeepSeek-V3 (checkpoint unspecified)82.2
7Llama 3.3 70B Instruct81.6#1385
8GPT-4.181.3#505
9Qwen 3.5 35B A3B80.9#492
10GPT-3.5 Turbo75.4#3005
11GPT-4o Mini75.2#1667
12Llama 3.1 8B Instruct72.1#4600

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

Interactive version: theaggregate.ai/benchmark?slug=graphinstruct-generation · How It Works · Data refreshed daily, snapshot 2026-11-21.