Researchers have released MiroThinker v1.0, an open-source research agent that utilizes a new "interactive scaling" approach to enhance tool-augmented reasoning and information-seeking capabilities. Unlike traditional methods that rely solely on increasing model size or context length, this method systematically trains the model to handle deeper and more frequent interactions with its environment.
- The 72B variant achieves up to 81.9% accuracy on GAIA, 37.7% on HLE, 47.1% on BrowseComp, and 55.6% on BrowseComp-ZH.
- With a 256K context window, the model can perform up to 600 tool calls per task through reinforcement learning.
- The approach leverages environment feedback to correct errors and refine trajectories, avoiding degradation risks associated with longer isolated reasoning chains.
The authors consider this significant because it establishes interaction scaling as a third critical dimension for building next-generation open research agents, complementing model capacity and context windows.