Research map / Reinforcement learning / RL from feedback
AlphaRouter: Token-level Routing Between SLM and LLM with Reinforcement Learning and Tree Search
Siteng Liao, Yuzhu Liang, Hengzhong Rao, Xizhao Luo, Tian Wang
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 TAPO-learned routing instead of routing to LLM at every predicted mismatch (ablation) | AlphaRouter · GSM8K (~1.0B active params) | accuracy (%): 64.50% -> 78.00% | Improved |
| add TAPO-learned routing instead of mismatch-triggered routing (ablation) | AlphaRouter · GSM8K (~1.0B active params) | LLM call rate (%): 8.80% -> 6.45% | Improved |
| add tree-advantage (sibling-branch) credit assignment instead of GRPO-style rollout objective (ablation) | AlphaRouter · GSM8K | accuracy (%) | Improved |
| increase CITS path budget M from 16 to 32 (ablation) | AlphaRouter · GSM8K | accuracy (%): 86.00% -> 90.50% | Improved |
| increase CITS expansion rounds N from 2 to 3 (M=32) (ablation) | AlphaRouter · GSM8K | accuracy (%): 89.50% -> 90.50% | Improved |
| transfer router trained on 0.5B+7B to unseen 3B LLM | AlphaRouter · GSM8K (0.5B+3B pair) | accuracy (%): 86.0% -> 90.0% | Improved |
| use AlphaRouter token routing instead of R2R at ~1.0B active params | AlphaRouter · GSM8K (~1.0B active params) | accuracy (%): 71.0% -> 78.0% | Improved |
| use AlphaRouter at 91% accuracy instead of standalone Qwen-7B decoding | AlphaRouter · GSM8K system efficiency | latency (s): 3.8420 -> 2.9459 | Improved |
| use AlphaRouter at 91% accuracy instead of standalone Qwen-7B | AlphaRouter · GSM8K system efficiency | computation (TFLOPs): 588.19 -> 259.92 | Improved |
| use AlphaRouter instead of speculative decoding at 91% accuracy | AlphaRouter · GSM8K system efficiency | computation (TFLOPs): 605.64 -> 259.92 | Improved |
Closest papers by meaning
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- TokenRouter: Efficient Serving System for Token-Level LLM Routing (NeurIPS 2026 · similarity 0.88)
- MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · similarity 0.87)
- R2-Router: A New Paradigm for LLM Routing with Reasoning (ICML 2026 · similarity 0.87)
- Bridging On-Device and Cloud LLMs for Collaborative Reasoning: A Unified Methodology for Local Routing and Post-Training (ICML 2026 · similarity 0.87)
- Causal LLM Routing: End-to-End Regret Minimization from Observational Data (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.
