Long research runs can fail because a model never learned how to investigate, even when its language skills are strong. The central insight here is to teach agent behavior before post-training, then refine it through environmental feedback rather than rely on elaborate orchestration alone.
Key Findings
- Agentic mid-training bridges ordinary pretraining and task-specific post-training, giving the model prior experience with planning, reasoning, and tool decisions.
- Automatically synthesized questions and trajectories support both supervised cold starts and reinforcement learning; stage-specific environments make costly web interaction more manageable.
- The reported model reaches 32.9 on Humanity's Last Exam and 90.6 on FRAMES, but its 43.4 BrowseComp score remains below OpenAI o3's 49.7 in the same comparison. Strength varies by task.
Who Should Read It
Great fit if you are designing training data, environments, or reinforcement learning for research agents. Look elsewhere if you need evidence of unrestricted tool generalization or consistently faithful long reports: the authors identify predefined tools, 128K context, and report fidelity as continuing constraints.