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

Title: AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use (Apr 2026)
Link: http://arxiv.org/abs/2604.21590v1
Date: April 2026

Summary:
This paper introduces AgenticQwen, a family of small agentic language models (8B and 30B) designed for efficient multi-step reasoning and industrial-scale tool use. The models are trained using multi-round reinforcement learning (RL) powered by a dual data flywheel framework: a reasoning flywheel that learns from errors to increase task difficulty and an agentic flywheel that evolves linear workflows into complex multi-branch behavior trees. Evaluation results demonstrate that these models significantly narrow the performance gap with much larger foundation models on benchmarks like TAU-2 and BFCL-V4 while maintaining lower inference costs.

Key Topics:
- Agentic Language Models
- Reinforcement Learning (RL)
- Dual Data Flywheels
- Synthetic Data Generation
- Tool Use
- Behavior Trees
- Knowledge Distillation

Chapters:
00:00 - Introduction and Industry Challenges
01:21 - Overcoming The Homogeneity Wall
03:01 - Reasoning Data Flywheel Explained
04:54 - Creating Contextual Diversity
06:24 - Implementing Consensus Filtering
07:46 - Dynamic Agentic Flywheel Overview
09:13 - Branch to Task Conversion
10:46 - Adversarial Mock User Scenarios
12:36 - Benchmarking Performance Results
14:23 - Industrial Use Case Evaluation
16:06 - Mixture Of Experts Efficiency
17:52 - Future Context Challenges

Stock video credits:
- Pressmaster - https://www.pexels.com/@pressmaster
- Google DeepMind - https://www.pexels.com/@googledeepmind
- Nino Souza - https://www.pexels.com/@ninosouza
- cottonbro studio - https://www.pexels.com/@cottonbro
- Bedrijfsfilmspecialist.nl - https://www.pexels.com/@bedrijfsfilmspecialist-nl-1284006
- Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
- tunnel motions - https://www.pexels.com/@tunnelmotions
- Max Fischer - https://www.pexels.com/@max-fischer
- Cyriac von Czapiewski - https://www.pexels.com/@cyriac-von-czapiewski-1601520
- Tima Miroshnichenko - https://www.pexels.com/@tima-miroshnichenko
- Kelly - https://www.pexels.com/@kelly
- Pachon in Motion - https://www.pexels.com/@pachon-in-motion-426015731
- Soumya - https://www.pexels.com/@soumya-1446957
- Anthony 🙂 - https://www.pexels.com/@inspiredimages
- Magda Ehlers - https://www.pexels.com/@magda-ehlers-pexels
- Vlada Karpovich - https://www.pexels.com/@vlada-karpovich
- KoolShooters - https://www.pexels.com/@koolshooters
- crazy motions - https://www.pexels.com/@crazy-motions-80195021
- Kindel Media - https://www.pexels.com/@kindelmedia
- Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
- Tom Fisk - https://www.pexels.com/@tomfisk
- Colin Jones - https://www.pexels.com/@larchmedia
- Adis Resic - https://www.pexels.com/@adis-resic-297996969
- fauxels - https://www.pexels.com/@fauxels
- Glenn Langhorst - https://www.pexels.com/@glenn-langhorst-120487
- Claudiu Ciobanu - https://www.pexels.com/@claudiuciobanu
- Anete Lusina - https://www.pexels.com/@anete-lusina
- Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
- Ketut Subiyanto - https://www.pexels.com/@ketut-subiyanto
- Ron Lach - https://www.pexels.com/@ron-lach
- Silviu Din - https://www.pexels.com/@silviu-din-1620549
- José Alfredo Munguía Lira - https://www.pexels.com/@rectorretro
- olia danilevich - https://www.pexels.com/@olia-danilevich
- Caleb Oquendo - https://www.pexels.com/@caleboquendo
- Oleg Gamulinskii - https://www.pexels.com/@oleg-gamulinskii-755060
- Colors Motion Graphics - https://www.pexels.com/@colors-motion-graphics-183847699
- Trippy Lagoon - https://www.pexels.com/@trippy-lagoon-511515544
- Dan Cristian Pădureț - https://www.pexels.com/@paduret
- Pixabay - https://www.pexels.com/@pixabay
- Thirdman - https://www.pexels.com/@thirdman

Видео AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale To канала AI Paper Slop