Designing Reasoning Systems for Mathematical Research

2 minute read

Published:

Design reasoning systems that help an agent solve real mathematical problems—or reach the same justified result with substantially less work.

The course focuses on birational geometry. Its lasting subject is how to specify a problem, organize proof search, use the literature, and check evidence. Models and tools will change; these mathematical responsibilities remain.

Outline of the posts

Follow this sequence while working on the same proof project.

  1. Represent the mathematics — keep the target, assumptions, and unresolved obligations precise.
  2. Plan and search — compare proof routes and choose the next useful task.
  3. Use the literature — retrieve the right theorem and justify its application.
  4. Preserve the implication back — close the gap between a reduction and the original problem.
  5. Review and admit results — locate proof gaps and determine what can be reused.
  6. Formalize and check — understand Lean, formal specifications, and the limits of a checked declaration.
  7. Remember and recover — preserve valid work when hypotheses or sessions change.
  8. Coordinate the work — divide tasks only when it improves the result or cost.
  9. Evaluate the system — compare capability and resource use at the same proof quality.

What counts as success?

An intervention should either close a mathematical gap the direct run could not close, or produce an equally sound result with less measured token use, time, repeated search, or human correction. Include preparation, retries, and added mathematical hints in the comparison. More agents or more elaborate records are not achievements by themselves.

For the final project, apply the workflow to a proof or proposed extension from your own research. Submit the actual argument, a precise account of remaining gaps, and a comparison with your baseline. A substantive partial result or a decisive obstruction can be valuable; an untested plan is insufficient.

The long-term aim is autonomous proof construction with mechanically checkable evidence. Formal checking establishes the encoded statement relative to its foundations; alignment with the intended question remains a separate requirement. This course does not assume that routine human proof checking must remain permanent.

References and evidence

Primary sources and claim boundaries

The source map distinguishes established mathematics, reported systems, and proposed course workflows. The workflows have not yet been validated by controlled agent runs; improvements must be demonstrated through actual results.