AdvancedMathBench Caps Top LLM at 75.8% on Undergrad Proofs

A new benchmark suite dropped on Hugging Face this week that is worth flagging if you care about how far frontier reasoning models actually reach on real mathematics. AdvancedMathBench pairs a 296-problem proof generation set called ProverBench with an 888-trajectory verification set called VerifierBench. The point is to look beyond final-answer accuracy and stress test whether models can construct valid proofs and, separately, judge proofs written by other models.

The numbers are the interesting part. On the proof generation side, the best model tested reaches only 75.8% on the undergraduate split and drops to 66.1% on the doctoral qualifying-exam split. On verification, the best balanced F1 across models tops out at 65.1%, with the authors flagging low true negative rates and critical-error detection as the standout weakness. In plain terms, models are notably better at drafting plausible proofs than at catching the flaws in each other's work.

Why this matters for anyone building on top of these models: proof-style reasoning shows up in more places than literal mathematics, from legal argument to code correctness to safety-critical planning. If verification is the weaker leg, the pattern of letting a model check its own work, which a lot of agent pipelines lean on, has a ceiling here worth understanding. It also gives a cleaner axis than olympiad-style contests, which have started to saturate at the top.

The honest caveat is that this is a fresh benchmark, and the leaderboard reflects a specific slate of models the authors chose to run. What the writeup does not give you is a public reproduction pipeline, a timeline for when a broader model set will be tested against the same rubric, or a breakdown of the subject-area distribution inside the 296 problems.

The forward-looking thing to watch is whether other labs start posting their models' numbers against ProverBench and VerifierBench. If they do, it becomes a more useful public thermometer for the next year of reasoning-model claims than another round of MATH or AIME scores.