Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2604.12385 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910128204152832 |
|---|---|
| author | Zhang, Jiarui Liu, Xiangyu Hu, Yong Niu, Chaoyue Zeng, Hang Tang, Shaojie Wu, Fan Chen, Guihai |
| author_facet | Zhang, Jiarui Liu, Xiangyu Hu, Yong Niu, Chaoyue Zeng, Hang Tang, Shaojie Wu, Fan Chen, Guihai |
| contents | Multi-turn dialogue is the predominant form of interaction with large language models (LLMs). While LLM routing is effective in single-turn settings, existing methods fail to maximize cumulative performance in multi-turn dialogue due to interaction dynamics and delayed rewards. To address this challenge, we move from myopic, single-turn selection to long-horizon sequential routing for multi-turn dialogue. Accordingly, we propose DialRouter, which first performs MCTS to explore dialogue branches induced by different LLM selections and collect trajectories with high cumulative rewards. DialRouter then learns a lightweight routing policy from search-derived data, augmented with retrieval-based future state approximation, enabling multi-turn routing without online search. Experiments on both open-domain and domain-specific dialogue tasks across diverse candidate sets of both open-source and closed-source LLMs demonstrate that DialRouter significantly outperforms single LLMs and existing routing baselines in task success rate, while achieving a superior performance-cost trade-off when combined with a cost-aware reward. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_12385 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | From Myopic Selection to Long-Horizon Awareness: Sequential LLM Routing for Multi-Turn Dialogue Zhang, Jiarui Liu, Xiangyu Hu, Yong Niu, Chaoyue Zeng, Hang Tang, Shaojie Wu, Fan Chen, Guihai Computation and Language Multi-turn dialogue is the predominant form of interaction with large language models (LLMs). While LLM routing is effective in single-turn settings, existing methods fail to maximize cumulative performance in multi-turn dialogue due to interaction dynamics and delayed rewards. To address this challenge, we move from myopic, single-turn selection to long-horizon sequential routing for multi-turn dialogue. Accordingly, we propose DialRouter, which first performs MCTS to explore dialogue branches induced by different LLM selections and collect trajectories with high cumulative rewards. DialRouter then learns a lightweight routing policy from search-derived data, augmented with retrieval-based future state approximation, enabling multi-turn routing without online search. Experiments on both open-domain and domain-specific dialogue tasks across diverse candidate sets of both open-source and closed-source LLMs demonstrate that DialRouter significantly outperforms single LLMs and existing routing baselines in task success rate, while achieving a superior performance-cost trade-off when combined with a cost-aware reward. |
| title | From Myopic Selection to Long-Horizon Awareness: Sequential LLM Routing for Multi-Turn Dialogue |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2604.12385 |