GraphInstruct (Generation) - L2 Multi-Constraint Composition: leaderboard

GraphInstruct (Generation) L2 Multi-Constraint Composition scores zero-shot graph generation on the 200 GraphInstruct (2026) instructions whose graphs must satisfy four or more joint constraints. The level quality score, 0 to 100 and higher is better, is computed by code from validity, distribution match, reference similarity and constraint satisfaction.

Metric: Level quality score (0-100; the paper's 0-1 score times 100) on the 200 GraphInstruct instructions of level L2, which requires four or more joint constraints; 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: Estimated already saturated. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.693.4#173
2Qwen 3.5 122B A10B91.1#346
3Claude Sonnet 4 (20250514)90#405
4Qwen 3.5 397B A17B88.8#264
5Qwen 3.5 35B A3B87.2#492
6Llama 3.3 70B Instruct86.5#1385
7GPT-4.186.5#505
8GPT-4o84.3#751
9GraphInstruct DeepSeek-V3 (checkpoint unspecified)82.3
10GPT-3.5 Turbo60.8#3005
11Llama 3.1 8B Instruct59.7#4600
12GPT-4o Mini58.5#1667

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-l2-multi-constraint-composition · How It Works · Data refreshed daily, snapshot 2026-11-21.