GT Bench (Graph Theory) - Hard: leaderboard
Graph Theory Bench (Xu et al., 2026) tests whether LLMs can carry out classical graph algorithms in language on synthetic graphs given as prose, structured text, adjacency matrices or adjacency lists. The Hard subset holds min-cost max-flow, cycle counting, spanning-tree counting and biconnected components, plus the other tasks on dense graphs; accuracy in percent, zero-shot at temperature 0.01, higher is better.
Metric: Accuracy (%; Hard subset: min-cost max-flow, cycle count, spanning-tree count and biconnected components always, other tasks on dense graphs; zero-shot vanilla prompting at temperature 0.01, one run, over the 10,000 evaluation instances of 24 classical graph problems posed in four input representations (natural language, structured language, adjacency matrix, adjacency list); accuracy against the algorithmically verified answer). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini (2025-01-31) | 75.09 | #486 |
| 2 | DeepSeek R1 | 64.1 | #467 |
| 3 | QwQ-32B | 38.46 | #909 |
| 4 | Phi-4 | 32.97 | #1920 |
| 5 | Llama 3.3 70B Instruct | 30.4 | #1353 |
| 6 | GPT-4o (2024-08-06) | 28.94 | #731 |
| 7 | GPT-4o Mini (2024-07-18) | 27.47 | #1512 |
| 8 | Llama 3.1 8B Instruct | 19.05 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gt-bench-graph-theory-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.