AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use

arXiv cs.CL / 4/24/2026

📰 NewsDeveloper Stack & InfrastructureTools & Practical UsageModels & Research

共有:

Key Points

The paper introduces the AgenticQwen model family, designed as small agentic language models for industrial tool use under tight latency and cost constraints.
Training uses multi-round reinforcement learning on a mix of synthetic data and limited open-source data, combining reasoning-focused RL with agentic RL.
It proposes dual “data flywheels” that automatically generate progressively harder tasks: a reasoning flywheel that learns from errors and an agentic flywheel that converts linear workflows into branching behavior trees.
The authors validate performance on public agent benchmarks and also test in an industrial agent system, reporting strong benchmark results and improved parity with much larger models on search and data analysis.
They provide model checkpoints and parts of the synthetic dataset on Hugging Face, along with data synthesis and RL training code and an integration into EasyDistill.

Abstract

Modern industrial applications increasingly demand language models that act as agents, capable of multi-step reasoning and tool use in real-world settings. These tasks are typically performed under strict cost and latency constraints, making small agentic models highly desirable. In this paper, we introduce the AgenticQwen family of models, trained via multi-round reinforcement learning (RL) on synthetic data and a limited amount of open-source data. Our training framework combines reasoning RL and agentic RL with dual data flywheels that automatically generate increasingly challenging tasks. The reasoning flywheel increases task difficulty by learning from errors, while the agentic flywheel expands linear workflows into multi-branch behavior trees that better reflect the decision complexity of real-world applications. We validate AgenticQwen on public benchmarks and in an industrial agent system. The models achieve strong performance on multiple agentic benchmarks, and in our industrial agent system, close the gap with much larger models on search and data analysis tasks. Model checkpoints and part of the synthetic data: https://huggingface.co/collections/alibaba-pai/agenticqwen. Data synthesis and RL training code: https://github.com/haruhi-sudo/data_synth_and_rl. The data synthesis pipeline is also integrated into EasyDistill: https://github.com/modelscope/easydistill.