It started with us evaluating them on our own university-math benchmarks: U-MATH for problem-solving and μ-MATH for judging solution correctness (see the HF leaderboard: toloka/u-math-leaderboard)
tl;dr: R1 sure is amazing, but what we find is that it lags behind in novelty adaptation and reliability: * performance drops when updating benchmarks with fresh unseen tasks (e.g. AIME 2024 -> 2025) * R1-o1 gap widens when evaluating niche subdomains (e.g. university-specific math instead of the more common Olympiad-style contests) * same with going into altogether unconventional domains (e.g. chess) or skills (e.g. judgment instead of problem-solving) * R1 also runs into failure modes way more often (e.g. making illegal chess moves or falling into endless generation loops)
Our point here is not to bash on DeepSeek — they've done exceptional work, R1 is a game-changer, and we have no intention to downplay that. R1's release is a perfect opportunity to study where all these models differ and gain understanding on how to move forward from here
Me and my team recently released two benchmarks on university-level math: U-MATH (for University-MATH) and μ-MATH (for Meta U-MATH).
We're working a lot on complex reasoning for LLMs, and we were in particular interested in evaluating university-curricula math skills — in topics such as differential calculus and linear algebra — for their wide applicability and practicality.
We noticed that available benchmarks at the time were either at or below high-school level, or mainly leaning towards Olympiad-style problems, or synthetically generated from a set of templates / seeds.
We wanted focus on university curricula and we wanted "organic" variety, so we created our own benchmark using problems sourced from actual teaching materials used in top US universities — that is how U-MATH came to be.
We also, and that is my primary focus in particular, are very eager on studying and improving evaluations themselves, since the standard llm-as-a-judge approach is known to be noisy and biased, but that often remains unaccounted for. So we then created a U-MATH-derived benchmark to do "meta-evaluations" — i.e. evaluate the evaluators — which allows to quantify their error-rates, study their behaviors and biases, and so on.