Research map / Large language models / Language, safety and interpretability
ICL-Router: In-Context Learned Model Representations for LLM Routing
Chenxu Wang, Hao Li, Yiqun Zhang, Linyao Chen, Jianhao Chen, Ping Jian, Qiaosheng Zhang, 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 |
|---|---|---|---|
| add query reconstruction training (stage 1) (ablation) | ICL-Router · In-distribution (5 benchmarks, average) | accuracy: 74.01 -> 76.30 | Improved |
| add query reconstruction training (stage 1) (ablation) | ICL-Router · OOD (5 benchmarks, average) | accuracy: 64.06 -> 66.47 | Improved |
| use a larger embedding model (0.3B to 8B) (ablation) | ICL-Router · In-distribution (5 benchmarks, average) | accuracy: 74.45 -> 76.30 | Improved |
| use a larger embedding model (0.3B to 8B) (ablation) | ICL-Router · OOD (5 benchmarks, average) | accuracy: 64.55 -> 66.47 | Improved |
| increase in-context exemplars from 100 to 500 (ablation) | ICL-Router · In-distribution (5 benchmarks, average) | accuracy: 74.94 -> 76.30 | Improved |
| increase in-context exemplars from 500 to 1000 (ablation) | ICL-Router · In-distribution (5 benchmarks, average) | accuracy: 76.30 -> 75.75 | Hurt |
| increase in-context exemplars from 500 to 1000 (ablation) | ICL-Router · OOD (5 benchmarks, average) | accuracy: 66.47 -> 65.71 | Hurt |
| add new LLMs to the pool without retraining the router | ICL-Router · In-distribution (5 benchmarks, average) | accuracy: 76.3 -> 79.9 | Improved |
| add new LLMs to the pool without retraining the router | ICL-Router · OOD (5 benchmarks, average) | accuracy: 66.4 -> 69.9 | Improved |
| add new LLMs to the pool | MODEL-SAT · In-distribution and OOD | accuracy | Mixed |
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
- RouteLLM: Learning to Route LLMs from Preference Data (ICLR 2025 · similarity 0.91)
- Dr.LLM: Dynamic Layer Routing in LLMs (ICLR 2026 · similarity 0.91)
- Universal Model Routing for Efficient LLM Inference (ICLR 2026 · similarity 0.90)
- Lookahead Routing for Large Language Models (NeurIPS 2025 · similarity 0.90)
- RouterDC: Query-Based Router by Dual Contrastive Learning for Assembling Large Language Models (NeurIPS 2024 · similarity 0.90)
- GraphRouter: A Graph-based Router for LLM Selections (ICLR 2025 · similarity 0.89)
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.
