TCAndon-Router
π Introduction
In multi-agent systems, the ability to select the appropriate agent(s) to handle a user query is a key determinant of overall system performance.
TCAndonRouter is a reasoning-centric multi-intent routing module whose primary role is to perform agent routing in multi-agent systems. Beyond agent routing, TCAndonRouter can be applied to any intent-routing scenario, including agent skill selection.
The main advantages of TCAndonRouter include:
- Designed specifically for real-world enterprise applications
- Supports dynamic onboarding of new agents (intents) New agents can be added simply by appending their descriptions, without retraining
- Provides transparent and interpretable routing decisions, improving explainability, robustness, and cross-domain generalization, and making post-deployment bad-case analysis easier
- Effectively resolves agent conflicts caused by overlapping responsibilities, leading to higher-quality final responses. When multiple agents are applicable, TCAndonRouter preserves all relevant agents. Each downstream agent generates its own response, and a Refining Agent subsequently merges these outputs into a single final answer
TCAndonRouter is trained using Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (DAPO), and achieves state-of-the-art performance on large-scale, real-world enterprise datasets, including HWU64, MINDS14, SGD, and the Tencent Cloud ITSM dataset(QCloud).
| Models | CLINC150 | HWU64 | MINDS14 | SGD | QCloud |
|---|---|---|---|---|---|
| GPT-5.1 | 93.84 | 85.59 | 95.59 | 73.90 | 92.80/93.06 |
| Claude-Sonnet-4.5 | 94.21 | 87.40 | 96.20 | 76.02 | 88.82/94.25 |
| DeepSeek-v3.1-terminus | 88.29 | 88.10 | 95.72 | 79.70 | 94.09/91.89 |
| ArcRouter | 62.98 | 69.33 | 91.79 | 65.59 | - |
| Qwen3-Embedding-4B | 57.21 | 54.27 | 94.12 | 37.02 | - |
| Qwen3-4B-Instruct-2507 | 70.12 | 80.29 | 90.08 | 58.74 | 82.23/79.44 |
| TCAndonRouter | 91.25 | 91.63 | 96.70 | 91.58 | 95.21/92.78 |
π§ How to use
Please refer to GitHub for code usage.
from transformers import AutoModelForCausalLM, AutoTokenizer
from prompt import router_prompt
from utils import load_config
tokenizer = AutoTokenizer.from_pretrained("tencent/TCAndon-Router")
model = AutoModelForCausalLM.from_pretrained("tencent/TCAndon-Router", device_map="auto")
agents = load_config('config/hwu64_config.xml')
query = "Can you recommend any pub in mg road"
prompt = router_prompt.format(agents=agents) + 'user:' + query
messages = [{"role": "user", "content": prompt}]
encoding = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=False,
return_tensors="pt"
)
outputs = model.generate(encoding.to(model.device), max_new_tokens=2048)
output_text = tokenizer.decode(outputs[0])
Generate Agent Descriptions
If you want to use TCAndonRouter on your own dataset, you need to provide agent descriptions. The required format is defined in config/xxx_config.xml.
You can generate agent descriptions using an LLM via generate_agent_desc.py, or write them manually.
python generate_agent_desc.py --dataset hwu64 --limit 50
π€ Citation
If you use TCAndonRouter in your work, please cite our paper:
@article{zhao2026TCAndonRouter,
title={TCAndonRouter: Adaptive Reasoning Router for Multi-Agent Collaboration},
author={Jiuzhou Zhao, Chunrong Chen, Chenqi Qiao, Lebin Zheng, Minqi Han, Yanchi Liu, Yongzhou Xu, Xiaochuan Xu, Min Zhang},
journal={arXiv preprint:2601.04544},
year={2026}
}
- Downloads last month
- 30
Model tree for tencent/TCAndon-Router
Base model
Qwen/Qwen3-4B-Instruct-2507