Research map / Large language models / Language, safety and interpretability
Selective Deferred Routing: Enabling Cost-Efficient Collaboration between Local SLMs and Remote LLMs
Qijun Miao, Zhixuan Fang
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 SLM hidden states instead of BERT text embeddings (prompt only) (ablation) | SDR MLP decider, Llama-3.1-8B · SQuAD | cost-performance AUC: 0.6013 -> 0.8978 | Improved |
| use SLM hidden states instead of BERT text embeddings (prompt only) (ablation) | SDR MLP decider, Llama-3.1-8B · MMLU | cost-performance AUC: 0.6047 -> 0.6509 | Improved |
| use SLM hidden states instead of BERT text embeddings (prompt only) (ablation) | SDR MLP decider, Llama-3.1-8B · GSM8K | cost-performance AUC: 0.5476 -> 0.6706 | Improved |
| use SLM hidden states instead of BERT text embeddings (prompt only) (ablation) | SDR MLP decider, Qwen-2.5-7B · SQuAD | cost-performance AUC: 0.6419 -> 0.9204 | Improved |
| add SLM answer to prompt input with text embeddings (ablation) | SDR (P,T), Qwen-2.5-7B · GSM8K | cost-performance AUC: 0.4975 -> 0.7123 | Improved |
| add SLM answer to prompt input with text embeddings (ablation) | SDR (P,T), Llama-3.1-8B · SQuAD | cost-performance AUC: 0.6013 -> 0.6590 | Improved |
| add SLM answer to prompt input with hidden states (ablation) | SDR (P,H), Llama-3.1-8B · MMLU | cost-performance AUC: 0.6509 -> 0.7008 | Improved |
| add SLM answer to prompt input with hidden states (ablation) | SDR (P,H), Llama-3.1-8B · SQuAD | cost-performance AUC: 0.8978 -> 0.8946 | No clear effect |
| add SLM answer to prompt input with hidden states (ablation) | SDR (P,H), Qwen-2.5-7B · SQuAD | cost-performance AUC: 0.9204 -> 0.9207 | No clear effect |
| add SLM answer to prompt input with hidden states (ablation) | SDR (P,H), Qwen-2.5-7B · GSM8K | cost-performance AUC: 0.7461 -> 0.7746 | Improved |
| use full SDR (transformer-block decider) instead of best baseline FrugalGPT | SDR, Llama-3.1-8B local · SQuAD (main results) | cost-performance AUC: 0.5769 -> 0.9165 | Improved |
| use SDR instead of FrugalGPT | SDR, Qwen-2.5-7B local · GSM8K (main results) | cost-performance AUC: 0.7516 -> 0.7993 | Improved |
Closest papers by meaning
- MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · similarity 0.89)
- BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute (ICML 2025 · similarity 0.88)
- Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing (ICLR 2024 · similarity 0.88)
- Dr.LLM: Dynamic Layer Routing in LLMs (ICLR 2026 · similarity 0.88)
- TokenRouter: Efficient Serving System for Token-Level LLM Routing (NeurIPS 2026 · similarity 0.87)
- Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving (NeurIPS 2025 · similarity 0.87)
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.
