Research map / Large language models / Agents, code and math
Beyond Gemini-3-Pro: Revisiting LLM Routing and Aggregation at Scale
Shengji Tang, Weihao Lin, Peng Ye, Jingqi Ye, Hao Li, Yiqun Zhang, Xiaosong Wang, Bo Zhang, Shuyue Hu, Tao Chen, LEI BAI, Wanli Ouyang
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 |
|---|---|---|---|
| add adaptive routing-aggregation switch (ARS) (ablation) | JiSi (QRR+SAS) · Component ablation (average over three domains) | average score: 71.09 -> 72.15 | Improved |
| add support-set-based aggregator selection (SAS) (ablation) | JiSi (QRR) · Component ablation (average over three domains) | average score: 69.68 -> 71.09 | Improved |
| add query-response mixed routing (QRR) (ablation) | JiSi · Component ablation (average over three domains) | average score: 67.83 -> 69.68 | Improved |
| add ARS alone without QRR and SAS (ablation) | baseline aggregation · Component ablation (average over three domains) | average score: 67.83 -> 64.87 | Hurt |
| add SAS alone (ablation) | baseline aggregation · Component ablation (average over three domains) | average score: 67.83 -> 68.74 | Improved |
| add response similarity to query similarity in support set (ablation) | JiSi routing · Similarity criterion | deviation (lower better): 34.40 -> 32.96 | Improved |
| add length (token cost) similarity (ablation) | JiSi routing · Similarity criterion | deviation (lower better): 32.96 -> 31.27 | Improved |
| apply adaptive switch to filter references on Hard questions (ablation) | JiSi · Reference accuracy, Hard | accuracy: 22.25 -> 30.74 | Improved |
| apply adaptive switch on Easy questions (ablation) | JiSi · Reference accuracy, Easy | accuracy: 95.37 -> 98.07 | Improved |
| aggregate with JiSi instead of routing only (query-response mixed router) (ablation) | JiSi · Nine-benchmark average (LLMRouterBench 30% test split) | average score: 69.68 -> 72.15 | Improved |
| scale majority voting from 3 best LLMs to 10 LLMs | Majority Voting · Nine-benchmark average (LLMRouterBench 30% test split) | average score: 64.64 -> 61.69 | Hurt |
| use four rollouts self-consistency instead of one best LLM | Best LLM · Nine-benchmark average (LLMRouterBench 30% test split) | average score: 70.55 -> 70.66 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RouterDC — RouterDC: Query-Based Router by Dual Contrastive Learning for Assembling Large Language Models (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.88)
- Dr.LLM: Dynamic Layer Routing in LLMs (ICLR 2026 · similarity 0.87)
- GraphRouter: A Graph-based Router for LLM Selections (ICLR 2025 · similarity 0.87)
- BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute (ICML 2025 · similarity 0.86)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.86)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
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.
