TCAndon-Router: Adaptive Reasoning Router for Multi-Agent Collaboration

AI summary

Overview

Research area: multi-agent systems (MAS), LLM routing, and reasoning-driven agent orchestration for enterprise applications.

Technical level: Advanced — the paper assumes familiarity with supervised fine-tuning, reinforcement learning (DAPO), reward shaping, and model merging (Slerp).

One-sentence scope: This paper introduces TCAndon-Router (TCAR), a 4B-parameter router that produces an explicit natural-language reasoning chain and then selects a subset of expert agents, whose outputs are fused by a Refining Agent.

What This Paper Is About

Enterprise multi-agent systems must decide which expert agent (or agents) should handle each user query, and existing "task routers" make a single, static label choice. That single-label assumption breaks down when agent responsibilities overlap or a query spans several domains, producing "agent conflicts" and brittle decisions. TCAR reframes routing as a "reason-then-select" problem: the router writes a reasoning chain, outputs all plausibly relevant agents, and a downstream Refining Agent merges their separate answers into one final response.

Key Contributions

1. A reasoning-centric multi-agent routing system. The authors describe TCAR as the first domain router that outputs natural-language decision rationales and supports flexible multi-agent selection, rather than a single hard label.
2. Conflict resolution through collaborative execution. A "multi-agent generation and refinement" pipeline is proposed, in which selected agents answer independently and a Refining Agent aggregates and reconciles their outputs.
3. Dynamic agent onboarding without retraining. Adding a new agent only requires appending its natural-language description to the existing agent set; the router itself is not modified or retrained.
4. An open-source collaborative routing framework. The model and associated resources are released at https://huggingface.co/tencent/TCAndon-Router.

Main Findings

• TCAR achieves state-of-the-art on four of five datasets. TCAR(4B) scores 91.25 on CLINC150, 91.63 on HWU64, 96.70 on MINDS14, 91.58 on SGD, and 93.98 (F1) on QCloud.
• CLINC150 is the exception. GPT-5.1 (93.84) and Claude-Sonnet-4.5 (94.21) both exceed TCAR's 91.25 there; the paper attributes this to CLINC150's 150 intent categories producing prompts that average 18k tokens, which challenges smaller models.
• Large gains on multi-turn and multilingual data. On SGD, TCAR scores 91.58 versus 79.70 for DeepSeek-v3.1-terminus, 76.02 for Claude-Sonnet-4.5, and 73.90 for GPT-5.1. On MINDS14, TCAR scores 96.70.
• TCAR exceeds leading general-purpose LLMs on QCloud. Its 93.98 F1 on the enterprise dataset is above DeepSeek-v3.1-terminus (92.98), GPT-5.1 (92.93), and Claude-Sonnet-4.5 (91.45).
• Baselines lag substantially. ArcRouter scores 62.98 (CLINC150), 69.33 (HWU64), 91.79 (MINDS14), 65.59 (SGD), and is not reported on QCloud. Qwen3-Embedding-4B scores 57.21, 54.27, 94.12, 37.02, and is not reported on QCloud. The untrained Qwen3-4B-Instruct-2507 base scores 70.12, 80.29, 90.08, 58.74, and 80.81.
• Reasoning chains help. Removing reasoning lowers scores to 89.65 (CLINC150), 87.83 (HWU64), 96.45 (MINDS14), 84.96 (SGD), 91.33 (QCloud); with reasoning they rise to 91.25, 91.63, 96.70, 91.58, 93.98.
• RL adds a large gain on SGD. SFT-only reaches 91.32, 90.30, 96.69, 86.30, 93.77; adding DAPO-based RL gives 91.25, 91.63, 96.70, 91.58, 93.98. The paper states RL improves recall while maintaining high precision and counteracts SFT-only conservatism of outputting a single agent.
• Slerp initialization does not improve final metrics but preserves exploration. SFT+Slerp yields 90.43, 91.45, 96.82, 91.53, 93.20 versus 91.25, 91.63, 96.70, 91.58, 93.97 for SFT-only initialization, while the Slerp-initialized model maintains higher entropy during RL.
• The Refining Agent improves response quality. Using DeepSeek-v3.1 as the underlying LLM, the Refining Agent achieves a significantly higher human preference win rate than randomly selecting a single agent, particularly for troubleshooting-type queries; the specific win-rate value is not reported in the provided text.
• Deployment cost stays modest. The average number of routed agents is 1.37, average reasoning length is under 100 tokens, average GPU latency is less than 1 second, and downstream execution adds an average of 0.37 additional agent calls.
• Three dominant error types. Incomplete problem descriptions, insufficient understanding of highly domain-specific scenarios, and mismatch between the reasoning chain and the final prediction.

Methodology in Plain English

The router receives a prompt built by concatenating a routing instruction, the user query, and natural-language descriptions of every available agent. Instead of emitting one label, it emits a <reason> block (deliberately not the <think> tag, to stay compatible with both instruction-following and think-style models) followed by one or more agent IDs; the released prompt allows up to 3 agents and supports an "oos" output when no agent fits.

Training happens in two stages. First, supervised fine-tuning on labeled data teaches the reason-then-select format using standard autoregressive loss. Second, reinforcement learning with DAPO refines agent selection. The reward has three parts: a precision-like term measuring the share of predicted agents that are correct, a coverage term measuring how much of the gold agent set is captured, and a length penalty that discourages listing too many agents. Samples where rollout consistency exceeds τ > 0.6 are removed from RL because they produce near-zero advantage variance. Results are reported on CLINC150, HWU64, MINDS14, SGD, and an internal Tencent Cloud dataset called QCloud, using accuracy for single-agent datasets and F1 for multi-agent datasets, plus an end-to-end task success rate.

At inference, all selected agents answer in parallel, and a Refining Agent compares, filters, and merges their responses into one answer — explicitly surfacing conflicts when the agents disagree.

Why This Matters

• Research impact. The work shifts task routing from single-label classification to reasoning plus subset prediction, and shows that conflicts can be preserved as multi-perspective evidence rather than compressed away. It also demonstrates dynamic agent addition without retraining, a practical requirement in growing deployments.
• Cloud troubleshooting. Multiple agents (network, CDN, EO) may be needed for a symptom like "webpage loads slowly"; the Refining Agent combines partial diagnoses into a complete resolution.
• Enterprise customer support and ITSM. The QCloud dataset covers networking, cloud storage, CDN, security, databases, and application operations, representing the high-ambiguity, high-conflict routing that enterprise help desks face.
• Multilingual and multi-turn assistants. Strong results on MINDS14 and SGD indicate applicability to cross-lingual routing and routed dialogue that spans several turns.
• Industry relevance. The work originates from Tencent Cloud Andon and targets real enterprise environments (finance, healthcare, education, cloud computing are named as MAS adopters), with the model open-sourced and deployable at 4B parameters on modest GPU requirements.

Future Directions

• Enforcing structured reasoning-chain constraints and improving alignment between the reasoning and the predicted agent set, since mismatches are a documented error type.
• Improving robustness to weak, ambiguous, or overly brief agent descriptions, which the authors identify as a critical dependency.
• Addressing long-tail and highly domain-specific knowledge (rare APIs, special network configurations) where training samples are sparse; the authors suggest fine-tuning on domain-specific private data when instruction formats change substantially.
• Designing more efficient collaboration protocols and low-cost expansion strategies for emerging business domains, to improve robustness and generalization.

Target Audience

ML engineers and applied researchers building production multi-agent or LLM-routing systems, especially those in cloud operations, ITSM, and enterprise customer support. It also suits readers interested in reinforcement learning for structured set prediction and in interpretable routing. Beginners will find the motivation accessible but the training methodology demanding, since it assumes familiarity with SFT, DAPO, reward shaping, and Slerp model merging.

Authors’ abstract

Multi-Agent Systems(MAS) have become a powerful paradigm for building high performance intelligent applications. Within these systems, the router responsible for determining which expert agents should handle a given query plays a crucial role in overall performance. Existing routing strategies generally fall into two categories: performance routing, which balances latency and cost across models of different sizes, and task routing, which assigns queries to domain-specific experts to improve accuracy. In real-world enterprise applications, task routing is more suitable; however, most existing approaches rely on static single-label decisions, which introduce two major limitations: (i) difficulty in seamlessly integrating new agents as business domains expand, and (ii) routing conflicts caused by overlapping agent capabilities, ultimately degrading accuracy and robustness.To address these challenges, we propose TCAndon-Router(TCAR): an adaptive reasoning router for multi-agent collaboration. Unlike traditional routers, TCAR supports dynamic agent onboarding and first generates a natural-language reasoning chain before predicting a set of candidate agents capable of handling the query. In addition, we design a collaborative execution pipeline in which selected agents independently produce responses, which are then aggregated and refined into a single high-quality response by a dedicated Refining Agent.Experiments on public datasets and real enterprise data demonstrate that TCAR significantly improves routing accuracy, reduces routing conflicts, and remains robust in ambiguous scenarios. We have released TCAR at https://huggingface.co/tencent/TCAndon-Router to support future research on explainable and collaborative multi-agent routing.