Apodex 1.1 将执行环境作为智能体扩展新维度
Apodex 1.1 将任务状态置于模型消息历史之外,把可执行环境本身作为扩展面,而非仅给智能体增加工具。该模型在 GDPVal、FrontierFinance 和 FrontierScience-Research 上分别提升 9.3、5.6 和 8.3 分。其训练环境包含真实文件、搜索和代码世界,具备状态转换、工具预算、失败条件和任务级验证器,使训练分布包含完整工作轨迹而非仅提示词与答案配对。
A context window is a terrible place to store the state of an 80-minute agent run.
One of the better ideas in Apodex 1.1 is that the task state sits outside the model's message history.
So @Apodex_AI is treating executable environments themselves as a scaling surface.
That is a much bigger idea than adding more tools to an agent.
So Apodex 1.1 Agent Team is simultaneously a trained general-purpose model and the larger execution stack used to turn that model into a long-running worker.
They reported gains of 9.3 points on GDPVal, 5.6 on FrontierFinance, and 8.3 on FrontierScience-Research.
Their training environments are actual file, search, and code worlds with state transitions, tool budgets, failure conditions, and task-level verifiers.
The model has to operate inside them, change the workspace, recover when something breaks, and eventually satisfy a delivery contract.
So the training distribution contains trajectories of work, not just prompts paired with good answers.
That distinction is so important for agents. You can train a model to know how a tool works and still have it fail the moment the authoritative file changes, a command errors halfway through, or two artifacts become inconsistent.
Apodex's argument is basically that the next scaling surface is the world the model learns to work inside.
在 X 查看被引用的帖子Our Apodex 1.1 technical report is now featured on @huggingface Daily Papers.
Complex work requires sustained interaction with files, information sources, and executable code—along with state maintenance, failure recovery, and verifiable delivery.
We call this working capability: sustained, verifiable progress toward a real-world objective.
Apodex 1.1 develops it through Environment Scaling and Agentic Coordination Scaling, connected by a shared execution harness and AgentOS.
Read and upvote: huggingface.co/papers/2608.23…
来源:Rohan Paul · x.com