Research map / Large language models / Language, safety and interpretability
The Avengers: A Routing Recipe for Collective Intelligence in Language Models
Yiqun Zhang, Hao Li, Chenxu Wang, Linyao Chen, Qiaosheng Zhang, Peng Ye, Shi Feng, Xinrun Wang, Jia Xu, Lei Bai, Shuyue Hu
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 Self-Consistency instead of Aggregation (ablation) | Avengers · Average excluding code tasks | average score: 56.73 -> 66.98 | Improved |
| use Self-Consistency instead of Direct (CoT) (ablation) | Avengers · Average excluding code tasks | average score: 59.19 -> 66.98 | Improved |
| use Self-Consistency instead of Model-Switch (ablation) | Avengers · Average excluding code tasks | average score: 66.42 -> 66.98 | No clear effect |
| swap embedding model (bge-m3 to text-embedding-3-large) (ablation) | Avengers · 15-dataset average | average score: 70.28 -> 70.70 | No clear effect |
| swap clustering algorithm (BIRCH to Hierarchical) (ablation) | Avengers · 15-dataset average | average score: 70.04 -> 70.71 | No clear effect |
| vary number of clusters K from about 14 to 140 (ablation) | Avengers · 15-dataset average | average score: surpasses GPT-4.1/4.5 for K approximately 14 to 140 | No clear effect |
| use only 30% of data for clustering/validation (ablation) | Avengers · 15-dataset average | average score: peaks at 0.30-0.35 test split; surpasses GPT-4.1/4.5 with 30% | No clear effect |
| route with clustering instead of mixture-of-agents | Avengers vs MoA · 15-dataset average | average score: 56.70 -> 70.54 | Improved |
| clustering-based router instead of EmbedLLM on OOD data | Avengers · 5 OOD datasets average | average score: 69.77 -> 74.42 | Improved |
| clustering-based router instead of EmbedLLM | Avengers · 15-dataset average | average score: 69.24 -> 70.54 | Improved |
| clustering-based router instead of Random Router | Avengers · 15-dataset average | average score: 49.78 -> 70.54 | Improved |
| clustering-based router instead of LLM Router | Avengers · 15-dataset average | average score: 64.76 -> 70.54 | Improved |
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 2 of 2 shared results
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
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.88)
- OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text (ICLR 2024 · similarity 0.87)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.87)
- Mixture-of-Agents Enhances Large Language Model Capabilities (ICLR 2025 · similarity 0.87)
- AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.86)
- MASS: Mathematical Data Selection via Skill Graphs for Pretraining Large Language Models (ICML 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.
