NetConfArena - Test-Case Score: leaderboard
NetConfArena (arXiv 2608.23179v1, Table II): LLM agents configure emulated router networks (static routing, RIP, OSPF, BGP, MPLS, VPN and overlays) in a closed loop over 480 task instances from parameterized templates. Mean share of executable diagnostic test cases the final network state satisfies, percent; thinking on and off are separate entries. Higher is better.
Metric: Mean fraction of task test cases satisfied (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro (Non-reasoning) | 96.1 | #112 (DeepSeek V4 Pro) |
| 2 | DeepSeek V4 Flash (Non-reasoning) | 91.4 | #155 (DeepSeek V4 Flash) |
| 3 | Qwen 3 32B (Thinking) | 72.8 | #449 (Qwen 3 32B) |
| 4 | Qwen 3 32B (Non-reasoning) | 65.6 | #449 (Qwen 3 32B) |
| 5 | Qwen 3 8B (Thinking) | 45.7 | #678 (Qwen 3 8B) |
| 6 | Qwen 3 8B (Non-reasoning) | 33.7 | #678 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=netconfarena-test-case-score · How It Works · Data refreshed daily, snapshot 2026-11-05.