Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem

2025/12/31 by Weixun Wang, XiaoXiao Xu, Wanhe An +86 · 22 voices · 2 citations
#cs.AI #cs.CL

paper · pdf

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 agentic model. 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, 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-Perceptive Agentic Policy Optimization (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 ALE.

Citations

• Group Sequence Policy Optimization
• VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks
• OpenHands: An Open Platform for AI Software Developers as Generalist Agents
• GAIA: a benchmark for General AI Assistants
• Efficient Memory Management for Large Language Model Serving with PagedAttention

Cited by

• SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration
• The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation
• Attention Illuminates LLM Reasoning: The Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Optimization

Discussions

• Worth reading this in full. I come in skeptical, but this basically is a claim that an AI system at Alibaba attempted autonomous replication without human intervention. This excerpt was found and high [bsky, 46 points, 3 comments]
• What arxiv.org/abs/2512.24873 [bsky, 30 points, 4 comments]
• Alibaba had their own version of "the agent is escaping the sandbox" back in December and almost nobody cared arxiv.org/abs/2512.24873 [bsky, 19 points, 1 comments]
• "Early one morning, our team was urgently convened after Alibaba Cloud’s managed firewall flagged a burst of security-policy violations originating from our training servers." The agent tried to mine [bsky, 15 points, 1 comments]
• Paper (found by Alexander Long): arxiv.org/pdf/2512.24873 [bsky, 9 points, 1 comments]
• A team working on a design for training AI models on workflows for software development found that it attempted to break free (reverse ssh out of its environment) and set up its own monetary supply (r [bsky, 7 points, 1 comments]
• The report detailing what occurred. arxiv.org/abs/2512.24873 [bsky, 5 points, 0 comments]
• Let It Flow: Agentic Crafting on Rock and Roll [hn, 3 points, 0 comments]
• AI agent broke out of lab to go learn how to mine cryptocurency arxiv.org/abs/2512.24873 [bsky, 3 points, 0 comments]
• un peligro real de los agentes de IA que debe ser tenido muy en cuenta a la hora de desplegarlos en entornos reales. Siempre, siempre, debe limitarse estrictamente los recursos a los que tienen acceso [bsky, 3 points, 1 comments]
• If Venezuela, Greenland, Cuba, Iran, and ufos were not enough to distract you from the Epstein files here is an AI deciding to crypto mine entirely by itself.Guess what will happen when someone asks i [bsky, 2 points, 0 comments]
• arxiv.org/pdf/2512.24873 [bsky, 2 points, 1 comments]
• Original report arxiv.org/abs/2512.24873 [bsky, 1 points, 0 comments]
• The paper introduces the Agentic Learning Ecosystem for improving AI agent development, featuring ROME, an agent trained on over a million trajectories, a novel policy optimization algorithm, and a pe [bsky, 1 points, 0 comments]
• Researchers discovered that ROME was trained on over a million trajectories using semantic chunks instead of tokens. It performs well on both SWE-bench Verified and Terminal Bench Pro. https://arxiv.o [bsky, 0 points, 0 comments]
• lol cryptocurrency is a virus even in managed environments https://arxiv.org/abs/2512.24873 [bsky, 0 points, 0 comments]
• 2/2 Let It Flow: Agentic Crafting on Rock and Roll Building the ROME Model within an Open Agentic Learning Ecosystem Page/Seite 15 arxiv.org/pdf/2512.24873 [bsky, 0 points, 0 comments]
• The really scary part is that it was a proof of concept that broke out of its sandbox. Here's the original paper. arxiv.org/pdf/2512.24873 [bsky, 0 points, 0 comments]
• AI that taught itself to misbehave [lemmy, 0 points, 0 comments]
• arxiv.org/abs/2512.24873 Did you read the full paper? What led you to the conclusion that it was fake? [bsky, 0 points, 0 comments]
• Now I'm not an expert, but I'm fairly sure that the second paragraph of 3.1.4 just casually offhandedly dropped that we've now created the infinite paperclip machine, as if there weren't enough reason [bsky, 0 points, 0 comments]
• arxiv.org/pdf/2512.24873 Alibaba agent mines crypto because why not wwxfromtju.github.io [bsky, 0 points, 0 comments]

Related

• Orchard: An Open-Source Agentic Modeling Framework
• Klear-AgentForge: Forging Agentic Intelligence through Posttraining Scaling
• AWorld: Orchestrating the Training Recipe for Agentic AI
• Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
• In-the-Flow Agentic System Optimization for Effective Planning and Tool Use