MiroThinker: Interactive Scaling for Research Agents

Translation status: The English translation is pending. The section structure is synchronized with the Chinese source.

Motivation: Interaction Depth as a Third Scaling Axis

Method: ReAct, Recency-Based Context Management, and Three-Stage Training

Agent Workflow

Data Synthesis and MiroVerse

Training Pipeline

Results

Evaluation Setup

Main Results

Interactive-Scaling Behavior

Limitations

Position in the Deep-Research Landscape