EnvFactory-8B

<h2 align="center"> EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL </h2> <p align="center"> <a href="https://arxiv.org/abs/2605.18703"> <img src="https://img.shields.io/badge/Paper-Arxiv-red?logo=arxiv&logoColor=red" alt="EnvFactory Paper on arXiv" /> </a> <a href="https://github.com/LARK-AI-Lab/EnvFactory"> <img src="https://img.shields.io/badge/GitHub-Code-181717?logo=github&logoColor=white" alt="GitHub Code" /> </a> <a href="https://lark-ai-lab.github.io/envfactory.github.io/"> <img src="https://img.shields.io/badge/GitHub-Page-4078c0?logo=github&logoColor=white" alt="GitHub Page" /> </a> <a href="https://huggingface.co/collections/LARK-Lab/envfactory"> <img src="https://img.shields.io/badge/Datasets-Hugging%20Face%20Data-orange?logo=huggingface&logoColor=yellow" alt="Datasets on Hugging Face" /> </a> <a href="https://huggingface.co/collections/LARK-Lab/envfactory"> <img src="https://img.shields.io/badge/EnvFactory-Hugging%20Face%20Model-FFCC00?logo=huggingface&logoColor=yellow" alt="EnvFactory on Hugging Face" /> </a> </p>

Overview

We propose EnvFactory, a fully automated framework that addresses the challenges of equipping LLMs with tool-use capabilities via Agentic Reinforcement Learning (Agentic RL). EnvFactory autonomously explores and verifies stateful, executable tool environments from authentic resources, and synthesizes natural multi-turn trajectories through topology-aware sampling and calibrated refinement, producing grounded queries with implicit intents.

This model is the official EnvFactory-8B trained from Qwen/Qwen3-8B using SFT and RL on synthesized tool-use trajectories.

Key Features

• Executable Environment Synthesis: Automatically discovers, validates, and deploys MCP-based tool environments from real-world APIs
• Topology-Aware Trajectory Sampling: Generates natural multi-turn tool-use trajectories that capture implicit human reasoning
• Robust RL Training: Uses verified environments and calibrated refinement for stable reinforcement learning
• Scalable Architecture: Achieves superior performance with significantly fewer environments (85 environments across 7 domains)

Training Details

Training Data

• SFT Data: LARK-Lab/EnvFactory-SFT-FILTERED - 53.4k filtered trajectories
• RL Data: LARK-Lab/EnvFactory-RL - 3.09k trajectories

Training Procedure

• SFT Stage: Full fine-tuning using LlamaFactory with DeepSpeed ZeRO-3
• RL Stage: Reinforcement learning using forked VeRL framework
• Base Model: Qwen/Qwen3-8B
• Training Epochs: 1 epoch for SFT
• Learning Rate: 1.0e-6 with cosine scheduler
• Batch Size: 1 per device with gradient accumulation of 32

Performance

Results on tool-use benchmarks compared to the base model:

Usage

Tool-Use Agent

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_path = "LARK-Lab/EnvFactory-8B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map="auto")

# Example tool-use conversation
messages = [
    {"role": "system", "content": "You are a helpful assistant with access to various tools."},
    {"role": "user", "content": "Search for recent papers about tool-use agents on arxiv."}
]

input_ids = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt").to(model.device)
outputs = model.generate(input_ids, max_new_tokens=1024, temperature=0.7, top_p=0.9)
response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

With MCP Tools

# Load MCP tool configuration
import json

with open("configs/mcp_server.json", "r") as f:
    mcp_config = json.load(f)

# Use with your preferred MCP client
# See https://github.com/LARK-AI-Lab/EnvFactory for integration details

Citation

If you find our work helpful, please consider citing:

@misc{xu2026envfactoryscalingtooluseagents,
      title={EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL}, 
      author={Minrui Xu and Zilin Wang and Mengyi DENG and Zhiwei Li and Zhicheng Yang and Xiao Zhu and Yinhong Liu and Boyu Zhu and Baiyu Huang and Chao Chen and Heyuan Deng and Fei Mi and Lifeng Shang and Xingshan Zeng and Zhijiang Guo},
      year={2026},
      eprint={2605.18703},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.18703}, 
}

License

This model is released under the Apache 2.0 License.