Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem
Weixun Wang, XiaoXiao Xu, Wanhe An, Fangwen Dai, Wei Gao, Yancheng He, Ju Huang, Qiang Ji, Hanqi Jin, Xiaoyang Li, Yang Li, Zhongwen Li, Shirong Lin, Jiashun Liu, Zenan Liu, Tao Luo, Dilxat Muhtar, Yuanbin Qu, Jiaqiang Shi, Qinghui Sun, Yingshui Tan, Hao Tang
2026-01-01
Summary
This paper introduces a new system called the Agentic Learning Ecosystem, or ALE, designed to make it easier to build and train AI agents that can interact with the real world and learn from their experiences.
What's the problem?
Currently, building these kinds of 'agentic' AI systems – ones that can take actions, see what happens, and improve over time – is difficult because there isn't a good, all-in-one set of tools available to the public. Researchers and developers often have to piece things together themselves, which is time-consuming and makes it harder to share and improve upon each other's work.
What's the solution?
The researchers created ALE, which has three main parts. First, ROLL helps fine-tune the AI model's 'brain' after its initial training. Second, ROCK provides a safe environment for the agent to practice and generate data. Finally, iFlow CLI is a tool to help manage the information the agent uses to make decisions. They also built an agent called ROME, trained using ALE on a huge amount of practice data, and a new set of tests, Terminal Bench Pro, to measure how well it performs. A key part of their training method, Interaction-based Policy Alignment, focuses on rewarding the agent for good *interactions* rather than just individual steps, making it learn more consistently over longer periods.
Why it matters?
This work is important because it provides the open-source community with a complete toolkit for building and studying agentic AI. By making it easier to create these agents, it can accelerate research and development in areas like robotics, automation, and personalized AI assistants, ultimately leading to more capable and helpful AI systems.
Abstract
Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. We introduce the Agentic Learning Ecosystem (ALE), a foundational infrastructure that optimizes the production pipeline for agent LLMs. ALE consists of three components: ROLL, a post-training framework for weight optimization; ROCK, a sandbox environment manager for trajectory generation; and iFlow CLI, an agent framework for efficient context engineering. We release ROME (ROME is Obviously an Agentic Model), an open-source agent grounded by ALE and trained on over one million trajectories. Our approach includes data composition protocols for synthesizing complex behaviors and a novel policy optimization algorithm, Interaction-based Policy Alignment (IPA), which assigns credit over semantic interaction chunks rather than individual tokens to improve long-horizon training stability. Empirically, we evaluate ROME within a structured setting and introduce Terminal Bench Pro, a benchmark with improved scale and contamination control. ROME demonstrates strong performance across benchmarks like SWE-bench Verified and Terminal Bench, proving the effectiveness of the ALE infrastructure.