Self-Revising Discovery Systems
From MIT, this paper argues that genuine scientific discovery is not answer generation but a change in the search space itself, and that an AI scientist must perceive that shift without being told. It develops a category-theoretic framework in which evidence, artifacts, operations, and verifiers are typed, and discovery is defined as a principled revision of that representational regime rather than more search within a fixed one.
Ask this paper
Discovery means changing the regime: The system is built to detect when the representational regime should change and to revise it autonomously. That reframes an AI scientist from a faster searcher into something that can move the boundaries of the space it searches.
A typed, categorical foundation: Evidence, artifacts, operations, and verifiers are formally typed. Old results are carried into the new regime by functorial transport, and what counts as genuine discovery is the residual content that transport alone cannot explain.
Description-length gates keep it honest: Proposed revisions are accepted only when they reduce total description length, which separates real structural gains from mere added complexity. In one run, 388 proposals yield just 25 accepted revisions, a deliberately strict 6.4% rate.
Why it matters: Two concrete instantiations, protein-mechanics modeling and a knowledge-computation graph with typed skills and validation checkpoints, show category theory serving as both a formal language and an engineering spec. It is a more principled blueprint for autonomous discovery than search-only AI scientists.