Research map / Large language models / Agents, code and math
RouterArena: An Open Platform for Comprehensive Comparison of LLM Routers
Yifan Lu, Rixin Liu, Jiayi Yuan, Xingqi Cui, Shenrun Zhang, Hongyi Liu, Jiarong Xing
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use commercial router GPT-5 with restricted OpenAI pool instead of open-source routers | GPT-5 · RouterArena overall | accuracy and cost: 74.0% accuracy at $14.02 per 1k queries, ranks #7 | Mixed |
| commercial NotDiamond vs open-source routers | NotDiamond · RouterArena overall | leaderboard rank: ranks #12; 68.0% at $9.34 | Hurt |
| use vLLM-SR and CARROT instead of always the strongest model | vLLM-SR, CARROT · normalized deferral curve | cost vs accuracy: roughly 35% lower cost with under 2% accuracy degradation | Improved |
| NIRT-BERT routing vs best single model | NIRT-BERT · normalized deferral curve | cost: baseline-level accuracy at 378% of the cost | Hurt |
| BERT-based routers under perturbed queries | vLLM-SR, MIRT-BERT, NIRT-BERT · routing robustness | routing consistency | Hurt |
| rely on OpenAI embedding API for routing | vLLM-SR, RouteLLM · routing latency | latency: other routers stay under 100ms | Hurt |
| raise router accuracy by deferring more to large models | binary routers such as RouteLLM · routing optimality | optimal selection ratio | Mixed |
| evaluate routers on hard queries rather than easy | all routers · RouterArena difficulty bands | accuracy: most routers above 89% on easy, often below 10% on hard | Hurt |
| use RouterDC with a pool of small models | RouterDC · RouterArena overall | accuracy and cost: lowest cost ratio, 32.0% accuracy | Mixed |
Closest papers by meaning
- R2-Router: A New Paradigm for LLM Routing with Reasoning (ICML 2026 · similarity 0.86)
- BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute (ICML 2025 · similarity 0.85)
- VL-RouterBench: A Benchmark for Vision-Language Model Routing (CVPR 2026 · similarity 0.85)
- Tackling the Data-Parallel Load Balancing Bottleneck in LLM Serving: Practical Online Routing at Scale (NeurIPS 2026 · similarity 0.85)
- TokenRouter: Efficient Serving System for Token-Level LLM Routing (NeurIPS 2026 · similarity 0.84)
- Dr.LLM: Dynamic Layer Routing in LLMs (ICLR 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
