When the Candidate Is a Search Strategy: Open Mathematics

AlphaEvolve’s mathematical experiments ask a different question from the matrix-multiplication case: can the system evolve a procedure that repeatedly finds better constructions across many open problems? This chapter focuses on that procedure, not only on the objects it eventually produces. The reader should come away able to recognize the staged heuristic-search pattern, interpret the reported success rates, and distinguish a verified construction or improved bound from a proof that an underlying optimum has been reached.

Why an open problem can still have a machine-checkable step

Many mathematical questions ask for an object with the best property known: a dense packing, a set with a favorable combinatorial relationship, a function satisfying an inequality, or a configuration with an improved geometric bound. The global optimum may be unknown, but a proposed object can often be checked locally. A program can construct the object, test whether it satisfies the required conditions, and calculate the quantity that determines how good it is.

That separation between local certification and global optimality is what makes these problems suitable for AlphaEvolve. The evaluator does not need to prove that no better construction exists in order to establish that a candidate is valid and better than a previous record. It can check the property and return a score. The score then tells the evolutionary search which directions deserve more attention. A new construction can therefore create real mathematical progress even when the open problem remains open.

The representation matters. In the configuration used for many of the mathematical experiments, AlphaEvolve evolves a heuristic search program rather than the final object directly. The heuristic receives a starting construction or the best construction found so far, searches for an improvement under a fixed time budget, and returns the best construction it manages to find. The program is judged by the quality of that output and by whether the construction satisfies the problem’s conditions.

This is a useful form of indirection. An object can be difficult to mutate as a static list of coordinates, intervals, or coefficients. A search procedure can instead express how to perturb, evaluate, and retain candidate objects. The procedure can also adapt to the current state. The same underlying problem may benefit from large exploratory moves when the construction is weak and from small, carefully targeted adjustments once the score is already close to the known record.

The construction lifecycle is the evolved artifact

The process begins with a simple or random construction. That choice is significant because the reported mathematical experiments did not rely on giving AlphaEvolve a near-solution for every problem. The initial object supplies a valid or at least usable state from which the first heuristic can search. The evaluator then measures the candidate’s objective and checks the relevant property.

A successful heuristic does not end the process. Its best construction becomes the starting point shown to another evolved heuristic. This creates a sequence of transitions: exploration from a weak state, evaluation of a candidate, retention of an improved best object, and another search that is conditioned on the new state. The program population evolves at the same time. One heuristic may be good at making large early gains; another may be better at polishing a configuration that is already strong.

The stages are not a fixed hand-written recipe.