R2ABench (Mini-SWE-Agent) - GED Accuracy: leaderboard
Requirements-to-architecture generation on R2ABench (2026): from a structured requirements specification of 68 educational and GitHub projects, the model driving the Mini-SWE-Agent framework writes a PlantUML architecture view, compared with a human-validated reference. Normalized graph-edit-distance accuracy between generated and reference graphs, over the views that parse, from 0 to 100; higher is better.
Metric: Normalized graph-edit-distance accuracy (0-100) between the generated and reference architecture graphs (L1), over the 68 R2ABench projects (17 educational and 51 GitHub repositories), generating a PlantUML architecture view from the structured requirements specification with the Mini-SWE-Agent framework under default settings and greedy decoding; scored only over the outputs that render and parse (the L0 gate), failed outputs excluded; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 4 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Coder 480B A35B | 51.76 | #709 |
| 2 | GPT-5 | 49.31 | #178 |
| 3 | Claude Sonnet 4.6 | 48.84 | #173 |
| 4 | DeepSeek V3.2 | 45.93 | #419 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=r2abench-mini-swe-agent-ged-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.