A visual map of the paper’s central claim: the step-change is not a better single answer, but a better mathematical workspace. These views focus on orchestration, persistent artifacts, parallel branches, reviewer loops, and benchmark behavior under long-horizon tool use.
One durable document fed by branching agents, tools, critiques, and reversible hypotheses.
The paper’s system shape is closer to a research group than a chatbot. Hover nodes to inspect roles.
Mock heatmap: where durable state accumulates across stages. Hotter cells mean more preserved evidence.
Step through a stylized search process: prove, disprove, compute, retrieve, and review branches evolve asynchronously.
Hover cells to see how the workbench shifts from ideation to verification instead of relying on one giant monologue.
Toggle between solve rate and test-time budget. The page uses mock comparators anchored to the transcript’s reported wins.
Three stylized stories from the transcript: where retrieval, computation, human correction, and review loops mattered.
Chat systems, formal provers, and the co-mathematician workbench score differently across workflow properties.
References grouped by orientation: formal proof, tool use, workflow, multi-agent search, and open-ended math research.