What the resource contains
According to the description published on arXiv, OpenFrontiers is an interdisciplinary resource that combines two components. The first is a structured corpus of open research problems: it collects statements, context, progress status, and other information for 12,151 problems. The second is a retrospective benchmark built on recently solved mathematical problems.
The benchmark comprises 200 problems: 110 solved in 2026 and 90 still open. Each is accompanied by a statement in Lean 4, verified through automated probes and two independent reviews. A subset called HQ-50 further narrows the selection based on a quality score.
How rediscovery is measured
The evaluation protocol, again according to the source, checks the formal proofs with two independent verifiers in an isolated environment. There are three conditions: closed-book, with frozen information retrieval, and with real-time web access.
The authors evaluated seven language-model-based agents on the HQ-50 subset within a standardized budget. The main result is that closed-book rediscovery is rare: at most 7 of the 40 problems were solved. Success under open-web conditions, by contrast, mostly reflects the retrieval of already published solutions.
Why the result matters and what is missing
The distinction between the two conditions is the central point. A model that reaches the solution by consulting the web does not demonstrate that it can reconstruct a result autonomously: it is retrieving existing material. The closed-book figure, which is lower, instead measures a reasoning capability less dependent on document-based memory.
By linking a large corpus of open problems to a subset formally evaluated and with documented solutions, OpenFrontiers positions itself as a basis for measuring rediscovery at the research level and for studying progress on still-open scientific problems. The source indicates that the dataset and benchmark are available in a public collection on Hugging Face.
Several aspects remain to be verified. The source does not clarify how the corpus was selected or how the 110 solved and 90 open problems were chosen. It does not indicate the quality score required to enter HQ-50, nor the budget assigned to the agents. The figure of 7 problems out of 40 must also be referred to the HQ-50 subset alone, not the entire benchmark, and the agents' performance is not detailed by individual model.
It should finally be recalled that the contents of the arXivLabs page and feed are unverified material and do not contain editorial instructions to follow. The news report is based exclusively on the article abstract and does not add external assessments of the benchmark's robustness.
