MIT Gives AI a Math Test to Prove It's Truly Discovering
MIT researchers built an AI scientist that uses category theory to mathematically separate genuine discovery from mere search, with measurable novelty.
Takeaways
- MIT's Buehler lab proposes a category-theoretic framework separating AI search from genuine discovery.
- True discovery is defined as a verified schema transition with measurable residual beyond Left Kan transport.
- Builder/Breaker system discovers mode-conditioned compliance law for protein flexibility under MDL gate.
- CategoryScienceClaw finds anisotropic stiffness rule for fiber networks via AIC gate, delta 123.87.
- Rejected models are preserved as typed provenance artifacts, not deleted from the discovery graph.
- Code released at github.com/lamm-mit/scienceclaw for the agentic substrate.
Most AI systems billed as autonomous scientists
are really sophisticated search engines. They explore inside a fixed vocabulary of variables, equations, and tools, but they cannot tell when the vocabulary itself needs to change. A new paper from MIT's Markus Buehler and Fiona Wang attacks this problem head-on, using category theory to give AI systems a formal way to know when they are doing real discovery versus just remixing what they already know.
The work, titled Self-Revising Discovery Systems for Science, lays out a mathematical framework where an AI achieves the ability to discover the scientific vocabulary it reasons in, with evidence, tools, artifacts, verifiers, failures and claims becoming typed provenance across three distinct modalities: retrieval (adding known objects), search (exploring a fixed schema), and discovery (a verified regime transition). The headline result: novelty stops being a vibe and becomes a number.
Why search is not discovery
The problem the authors set out to fix is structural. Current AI scientists are extremely good at recombining and optimizing inside a fixed scientific representation, but real scientific progress often comes from changing that representation entirely: a new variable, a new operation, a new verifier. When ProtAgents proposes a new protein or an AI Scientist drafts a paper, are they searching inside a fixed regime, or actually expanding the regime? Until now, there was no rigorous way to tell.
This is also why open-endedness
evaluation has been so hard. Benchmarks measure performance against a fixed test set, but a system that genuinely discovers something new should, by definition, break the benchmark.
The category theory toolkit
The framework treats an agentic discovery system as a typed artifact population over a schema category. In a fixed regime with schema category S_b, the system state is a copresheaf I_t mapping the schema to Set, and fixed-regime operation is an update on such states, endofunctorial only when provenance-preserving refinements are specified and preserved. In plain terms: every artifact (a sequence, a contact graph, a hypothesis, a model) has a declared type, every tool has a declared signature, and the entire research process becomes an auditable graph of typed objects and morphisms.