Tasks & learning environments
We turn real business problems into tasks agents can attempt and learn from. Starting with Chinese corporate information, we study data boundaries, task definitions, and evaluation methods as we work toward learning environments for complex workflows.
How do we turn complex real-world work into tasks agents can attempt, learn from, and be compared on?
This direction evolved from Mingcha’s company-research project. Information in real investigations becomes available at different times and does not always connect directly. We study task inputs, temporal boundaries, and answer requirements so evaluation reflects an agent’s ability to work under those conditions.
Define tasks from real questions.
Organize the questions, data, and processes of corporate research into clear tasks. Investigate which inputs and feedback reflect useful capabilities.
Bring time and data boundaries into evaluation.
Study when information becomes available, how sources connect, and which conditions an answer depends on, grounding evaluation in explicit data boundaries.
Work toward interactive environments.
Build on offline tasks and evaluation methods to explore tool interactions, stepwise feedback, and learning environments where agents can learn complex work through action.