AgentOmnia Framework for Full-Scenario Agentic AI Scaling
A new framework named AgentOmnia has been unveiled by researchers to enhance the scalability of agentic large language models for various applications, including To-Consumer, To-Business, and To-Employee sectors. This framework orchestrates the definition of tasks, data generation, post-training processes, assessment, and enhancements. It employs a flexible taxonomy of Domain x Capability x Atomic Difficulty, which is compatible with OmniaBench for detailed diagnostics. AgentOmnia integrates bidirectional environment-task synthesis with dependencies on tools, structured programs, and solver-based pipelines, resulting in the creation of 5,018 stateful environments, 255,375 tools, and 52,361 tasks. The post-training phase utilizes supervised fine-tuning, online agentic reinforcement learning, and a rollback curriculum, while evaluation emphasizes correctness signals from various sources. This initiative tackles fragmentation in domains, capabilities, task complexities, and interaction contexts.
Key facts
- AgentOmnia is a framework for scaling agentic LLMs across full-scenario applications.
- It covers To-Consumer, To-Business, and To-Employee domains.
- Uses Domain x Capability x Atomic Difficulty taxonomy with OmniaBench.
- Constructs 5,018 stateful environments with 255,375 tools and 52,361 tasks.
- Post-training includes supervised fine-tuning, online agentic RL, and rollback curriculum.
- Evaluation uses correctness signals from programs, solvers, and verifiers.
- Addresses fragmentation across domains, capabilities, task difficulty, and interaction settings.
- Published on arXiv with ID 2607.23124.
Entities
Institutions
- arXiv