EvoRouter - Reasoning: leaderboard
EvoRouter v1 compares multi-agent routing configurations with a fixed resource pool. It uses 1300 in-domain test questions across 5 domains, plus 175 LiveCodeBench and 200 BBH-LD out-of-domain questions. Results average 3 runs; the mean correct count is rounded before denominator normalization. Pool models use temperature 1 and an output budget of 1024 tokens (4096 for code). Nonvision models receive an explicit image-invisibility placeholder on multimodal tasks. Router 4B and 8B are trained Qwen3 variants; LLaMa 8B is a trained Llama3.1 variant. Systems without model selection use GLM-5. Distinct trained-model, router, and harness identities are preserved. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | EvoRouter study EvoRouter-8B | 96.5 | |
| 2 | EvoRouter study EvoRouter-LLaMa-8B | 96 | |
| 3 | EvoRouter study EvoRouter-4B | 94 | |
| 4 | EvoRouter study MAS-Orchestra | 86.5 | |
| 5 | EvoRouter study RCR-Router | 86.5 | |
| 6 | EvoRouter study MasRouter | 84.5 | |
| 7 | EvoRouter study xRouter | 81 | |
| 8 | EvoRouter study AgentRouter | 79 | |
| 9 | EvoRouter study DAAO | 76.5 | |
| 10 | EvoRouter study Qwen3.5-397B-A17B | 75 |
Interactive version: theaggregate.ai/benchmark?slug=evorouter-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-28.