CoVe: Training Interactive Tool-Use Agents via Constraint-Guided Verification

Jinpeng Chen; Cheng Gong; Hanbo Li; Ziru Liu; Zichen Tian; Xinyu Fu; Shi Wu; Chenyang Zhang

cs.AIMar 2, 2026

CoVe: Training Interactive Tool-Use Agents via Constraint-Guided Verification

Authors: Jinpeng Chen, Cheng Gong, Hanbo Li, Ziru Liu, Zichen Tian, Xinyu Fu, Shi Wu, Chenyang Zhang,

Organizations: Huawei Research · Independent Researcher

Abstract

Developing multi-turn interactive tool-use agents is challenging because real-world user needs are often complex and ambiguous, yet agents must execute deterministic actions to satisfy them. To address this gap, we introduce CoVe (Constraint-Verification), a post-training data synthesis framework designed for training interactive tool-use agents while ensuring both data complexity and correctness. CoVe begins by defining explicit task constraints, which serve a dual role: they guide the generation of complex trajectories and act as deterministic verifiers for assessing trajectory quality. This enables the creation of high-quality training trajectories for supervised fine-tuning (SFT) and the derivation of accurate reward signals for reinforcement learning (RL). Our evaluation on the challenging $τ^2$ -bench benchmark demonstrates the effectiveness of the framework. Notably, our compact CoVe-4B model achieves success rates of 43.0% and 59.4% in the Airline and Retail domains, respectively; its overall performance significantly outperforms strong baselines of similar scale and remains competitive with models up to $17\times$ its size. These results indicate that CoVe provides an effective and efficient pathway for synthesizing training data for state-of-the-art interactive tool-use agents. To support future research, we open-source our code, trained model, and the full set of 12K high-quality trajectories used for training.

Figures & tables

Figure 1: Performance evaluation on the $\tau^{2}$ -bench Airline and Retail domains. We report success rates across $pass^{1}$ to $pass^{4}$ metrics to assess both peak performance and stability.

Figure 2: The CoVe framework. Explicit constraints are sampled and fuzzified to guide a User Simulator LLM in generating ambiguous, realistic queries. Upon conversation completion, the original deterministic constraints act as a checklist to automatically verify the agent’s tool invocations.

Figure 3: The system prompt (top) and the first user prompt (bottom) for the User Simulator. After receiving the first user prompt, the simulator begins issuing task requests to the Agent. In subsequent dialogue turns, the Agent’s response serves as the user message for the User Simulator.

Proprietary Models
Model	Airline				Retail				Average
Model	$pass^{1}$	$pass^{2}$	$pass^{3}$	$pass^{4}$	$pass^{1}$	$pass^{2}$	$pass^{3}$	$pass^{4}$	$pass^{1}$	$pass^{2}$	$pass^{3}$	$pass^{4}$
GPT-5	62.5%	55.3%	51.0%	48.0%	81.6%	71.8%	64.7%	58.8%	72.1%	63.6%	57.9%	53.4%
GPT-4o	47.6%	35.7%	30.0%	26.5%	64.0%	50.8%	43.0%	37.7%	55.8%	43.3%	36.5%	32.1%
Open-Source Models ( $\geq$ 70B)
Qwen3-235B-A22B-Inst.	48.5%	37.7%	31.0%	26.0%	63.6%	51.2%	44.5%	39.5%	56.1%	44.5%	37.8%	32.8%

Table 1: Main results evaluating proprietary and open-source models on the Airline and Retail domains of

\tau^{2}

-bench. Bold and underline denote the best and second-best results within each group.

Training Set	Airline	Retail	Average
None	25.0%	40.1%	32.6%
APIGen-MT-5K	34.0%	49.3%	41.7%
Simia-5K	33.0%	46.3%	39.7%
Simia-90K	42.0%	46.5%	44.3%
CoVe-5K (Ours)	38.8%	50.5%	44.7%
CoVe-12K (Ours)	43.0%	59.4%	51.2%

Table 2: Performance comparison (

pass^{1}

) for the Qwen3-4B-Instruct-2507 model fine-tuned on different SFT datasets. Bold and underline denote the best and second-best results.

Training Set	Airline	Retail	Average
Qwen3-4B-Inst.	25.0%	40.1%	32.6%
SFT	43.0%	59.4%	51.2%
RL	32.5%	48.9%	40.7%
SFT+RL	38.0%	55.7%	46.9%

Table 3: Left: Performance comparison (

pass^{1}

) of different training paradigms using the CoVe framework on the Qwen3-4B-Instruct-2507 base model. Bold denotes the best results. Right: Success rates of SFT data generation across different domains using various User Simulator LLMs.

Tool-Using Agents

Tool-using agents plan and act through external tools, APIs, software interfaces, and environments to complete tasks beyond language generation.

See all 286 papers

Controllable and Verifiable Tool-Use Data Synthesis for Agentic Reinforcement Learning
Apr 10, 2026Siyuan Xu, Shiyang Li, Xin Liu +9Tool Use RL for Tool Use
The Pennsylvania State University · Amazon · Georgia Institute of Technology
PACT: Privileged Trace Co-Training for Multi-Turn Tool-Use Agents
Jun 15, 2026Zhenbang Du, Jun Luo, Zhiwei Zheng +8Learning from Demonstration LLM Agent Training
Georgia Institute of Technology · Ohio State University · University of Pennsylvania +1
TopoCurate:Modeling Interaction Topology for Tool-Use Agent Training
Mar 2, 2026Jinluan Yang, Yuxin Liu, Zhengyu Chen +7LLM Agent Training RL for Tool Use
Zhejiang University, Hangzhou, China · Meituan, Beijing, China.

Apr 10, 2026J/K move · Enter open · S save

Controllable and Verifiable Tool-Use Data Synthesis for Agentic Reinforcement Learning

Siyuan Xu, Shiyang Li, Xin Liu, Tianyi Liu+8

The Pennsylvania State University · Amazon · Georgia Institute of Technology

Tool Use RL for Tool Use Agentic RL Reinforcement Learning with Verifiable Rewards

Open paper AskPDF arXiv

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

Abstract

Tool-Using Agents

Explore similar work

Controllable and Verifiable Tool-Use Data Synthesis for Agentic Reinforcement Learning

PACT: Privileged Trace Co-Training for Multi-Turn Tool-Use Agents

TopoCurate:Modeling Interaction Topology for Tool-Use Agent Training