Loading page…
The project would connect ideas, arguments and competing approaches—not just produce answers. A preliminary version is expected for testing in a few months.
Listen to this story
On October 7, 2026, Caltech described a joint effort with the American Institute of Mathematics to build a research workspace that tracks conjectures, arguments, references and competing approaches instead of merely generating solutions. A test version remains months away; AIM’s more than 1,000 annual visiting mathematicians are expected to help evaluate it. The team plans to use a year of usage data to guide a major redesign, potentially including a new AI system, but says that phase depends on major backing. The practical test is whether experts find the tool useful in collaborative research.
The platform is intended to suggest connections and strategies while supporting several lines of reasoning in parallel.
A planned Lean integration would turn human mathematical reasoning into formal code; the team says current large language models do this inefficiently.
Yisong Yue says researchers’ “research taste” remains something AI lacks, and argues that understanding how results are reached matters.
Caltech and the American Institute of Mathematics are developing an AI companion meant to make mathematical research easier to follow, not just generate solutions. Caltech described the collaboration on October 7, 2026. Its ambitions include connecting ideas and tracking competing arguments, but a preliminary version is still a few months away from testing by mathematicians.
The effort joins Caltech’s Information Science and Technology initiative with AIM. Project leaders Sergei Gukov, Yisong Yue and Pietro Perona began working on the platform about nine months ago. Their goal is to combine mathematicians’ judgment about worthwhile problems with AI assistance that organizes the work along the way.
The proposed platform would show how conjectures—mathematical statements awaiting proof—connect with arguments, references and alternative approaches over time. Different researchers or small groups could pursue competing routes while the system connects their progress. They could revisit earlier ideas without losing sight of the larger project.
One planned integration is conversion of human mathematical reasoning into Lean code, a precise language for expressing mathematics to computers. The team says current large language models are inefficient at that conversion and intends to build it directly into the system.
Yue calls the judgment experts develop through experience “research taste,” something he says AI currently lacks. In an October 5 Washington Post interview with Benjamin Guggenheim, Gukov also described a proposal to train AI on the “invisible work” involved in solving complicated theorems.
AIM already welcomes more than a thousand visiting mathematicians each year for week-long sessions. Those visitors will be invited to use the platform and provide feedback alongside Caltech AI engineers. The team plans to refine the software continually over months and years rather than treat the preliminary version as a finished tool.
Gukov expects a year of usage data to inform a major redesign, including training a new AI system to support researchers’ work. That stage still needs support: he says the team is seeking major backers, while citing initial commitments from Paul Stahura and Rahim Noorani.
Loading discussion...
Join the conversation
Explain which would make an AI research companion more valuable.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.