| --- |
| license: apache-2.0 |
| pipeline_tag: text-generation |
| arxiv: 2512.24873 |
| tags: |
| - agent |
| - moe |
| --- |
| |
| # ROME-30B-A3B |
|
|
| <p align="left" style="display: flex; gap: 8px; align-items: center;"> |
| <a href="https://arxiv.org/pdf/2512.24873" target="_blank"> |
| <img src="https://img.shields.io/badge/Paper-arXiv%3A2512.24873-red" alt="Paper"> |
| </a> |
| <a href="https://faithful-almanac-add.notion.site/The-Bitter-Lesson-Behind-Building-Agentic-RL-in-Terminal-Environments-2eaddd45837f80c9ad2ed6a15ef3c1a1?pvs=74" target="_blank"> |
| <img src="https://img.shields.io/badge/Blog-Notion-orange" alt="Blog"> |
| </a> |
| <img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg" alt="License"> |
| <img src="https://img.shields.io/badge/Model%20Type-MoE-green" alt="Model Type"> |
| </p> |
| |
|
|
|
|
| --- |
|
|
|
|
| **ROME** (**R**OME is **O**bviously an **A**gentic **M**odEl) is an open-source **agentic model** incubated within the **ALE (Agentic Learning Ecosystem)**. |
|
|
| Rather than scaling performance purely by increasing parameter count, ROME achieves parameter-scale–crossing performance through full-stack infrastructure integration and advanced Reinforcement Learning optimization. |
|
|
| <img src="https://rlhf.oss-cn-hangzhou.aliyuncs.com/iFLOW-ROME/performance.png" width="600"/> |
|
|
|
|
| --- |
|
|
|
|
|
|
| ## 🚀 Highlights |
|
|
| <img src="https://rlhf.oss-cn-hangzhou.aliyuncs.com/iFLOW-ROME/ALE.PNG" width="600"/> |
|
|
|
|
| ### 🔧 ALE Full-Stack Infrastructure |
| - [**ROLL**](https://github.com/alibaba/ROLL) – Large-scale reinforcement learning optimization engine |
|
|
| - [**ROCK**](https://github.com/alibaba/ROCK) – Secure sandbox and environment orchestration for agent execution |
|
|
| - **iFlow CLI** – Unified agent framework and developer interface |
|
|
| |
|
|
| ### 🧠 IPA Policy Optimization Algorithm |
| - Introduces **Interaction-Perceptive Agentic Policy Optimization (IPA)** |
| - Performs credit assignment at the level of **Semantic Interaction Chunks** |
| - Significantly improves **training stability** and **success rates** on **long-horizon tasks** |
|
|
|
|
|
|
| ### 🚀 Strong Agentic Performance |
| - Despite being a **mid-sized model** (30B MoE with 3B active parameters), ROME outperforms same-scale models on standard agent benchmarks: |
| - **Terminal-Bench 2.0**: 24.72% |
| - **SWE-bench Verified**: 57.40% |
| |
| - Performance is competitive with, and in some cases comparable to, models exceeding **100B parameters** |
|
|
| |
|
|
| ### 🔒 Production-Grade Safety |
| - Designed for autonomous agent execution in real environments |
| - Rigorously aligned and red-teamed against risks such as: |
| - Unauthorized access |
| - Illegal or unsafe tool invocation |
| - Built with **deployment-grade safety guarantees** in mind |
|
|
| --- |
|
|
|
|
|
|
| ## 📊 Performance (Preview) |
|
|
| ### Terminal-Based Benchmarks |
|
|
| | **Model** | **Terminal-Bench 2.0** | **SWE-bench Verified** | |
| | ---------------------------- | ---------------------- | ---------------------- | |
| | Qwen3-Coder-30B-A3B-Instruct | 13.48% | 46.33% | |
| | **ROME-30B-A3B** | **24.72%** | **57.40%** | |
| | GPT-OSS-120B | 21.12% | 43.93% | |
| | GLM-4.5 Air (106B) | 17.30% | 56.20% | |
|
|
| > See the technical report for full experimental details. |
|
|
| --- |
|
|
|
|
|
|
| ## 📜 Citation |
|
|
| If you find our work useful, please consider citing: |
|
|
| ```bibtex |
| @article{rome2025ale, |
| title={Let It Flow: Agentic Crafting on Rock and Roll - Building the ROME Model within an Open Agentic Learning Ecosystem}, |
| author={Wang, Weixun and Xu, XiaoXiao and An, Wanhe and Dai, Fangwen and others}, |
| journal={arXiv preprint arXiv:2512.24873}, |
| year={2025} |
| } |
| ``` |
