HMMT

HMMT is a competition Harvard and MIT undergraduates have run twice a year since 1998, mostly for their own high-school-age competitors rather than for AI researchers. Its individual and guts rounds set 50-minute problems with short, checkable answers — a number or an algebraic expression rather than a multiple-choice option — which is exactly the property that let it be repurposed as a maths-reasoning eval once reasoning models needed something harder than AIME. February’s team round asks for full written proofs and is graded by hand, so it stays outside the automated benchmark; the short-answer rounds do not.

HMMT entered the model-release cycle alongside AIME through ETH Zurich’s MathArena, which extracts each sitting’s problems into a verified, machine-gradable set within days of the exam. Progress on it has tracked the same reasoning-model race: a single May 2025 point release of DeepSeek’s R1 lifted its HMMT 2025 score from 41.7% to 79.4% by letting the model reason for longer, and by mid-2026 Thinking Machines Lab reported 86.3% for its Inkling model and 90.2% for the smaller, distilled Inkling-Small on the same competition.

Because HMMT shares AIME’s basic vulnerability — problems and solutions circulate publicly soon after each sitting — MathArena scores models against the newest exam available and treats older sittings as increasingly unreliable once enough time has passed for them to reach training data, an approach that keeps HMMT usable without pretending contamination is solved rather than deferred.