OpenFrontiers
Join
About

A working index, not a museum

There is no shortage of unsolved problems. What is missing is a place where they are written down precisely, attributed honestly, kept current, and left open to anyone who has something to add — including a partial result, a dead end, or a machine transcript.

The paper

OpenFrontiers is the community edition of the corpus described in OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery (Jiarui Yao, Ruida Wang, Tong Zhang, 2026). The paper assembles 12,151 problem records from 62 source collections, keeps uncertain status as unassessed rather than promoting it to open, and builds OpenFrontiers-Bench — 200 Lean-checked mathematics problems, 110 resolved in 2026 and 90 still open — on top of it. The site starts from that snapshot and lets the community keep it current.

BibTeX
@article{yao2026openfrontiers,
  title   = {OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery},
  author  = {Jiarui Yao and Ruida Wang and Tong Zhang},
  journal = {arXiv preprint arXiv:2610.37739},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.37739}
}

What an entry is

Each entry carries a stable identifier (OP-0001 and upward), a statement written so that a specialist in a neighbouring field can read it, the provenance of where it was posed, and a section on what is already known. Browse by field, source and research status, then follow the original references for the precise scope of each question.

Resolved problems stay in the index, marked as resolved and with the resolution recorded. A problem that was open for forty years and fell last spring is more instructive than one that is merely open, and deleting it would throw away the most interesting part.

What it is not

  • —Not a journal. Nothing here is peer review in the formal sense; accepted submissions have been read by an editor, not refereed.
  • —Not an authority on status. If an entry says a problem is open and it was settled last month, that is a bug — please say so in the discussion and it will be fixed.
  • —Not a place to publish a claimed proof of the Riemann hypothesis without an argument. Submissions of that kind are declined without discussion.

Credits, quotas and machine attempts

Reading is free and needs no account. Everything expensive — running a model against a problem, writing through the API — is paid for in credits, and credits come from contributing. The intent is straightforward: the people who curate the corpus are the people who get to compute against it.

Opening balance on a new account+250
Monthly refresh+250
Problem published after review+400
Submission accepted+800
Reviewing for an editor+60
Machine attempt−25
API write call−1
API read callfree

Rate limits are 30 requests per minute without a key, 120 with one, and 600 for accounts with accepted contributions. A failed machine attempt is refunded.

On machine attempts

Models are now good enough at mathematics to be worth pointing at open problems, and bad enough that most attempts fail. Both facts argue for recording the transcripts. An attempt is stored in full, with the model’s own assessment of what it did and did not establish, and a verdict label. Attempts that claim a solution are flagged for a human, not published as results.

The failures are the point. A public record of which approaches have already been tried and where they stop is the thing the literature does not provide, for humans or machines.

Corrections

Mis-stated problems, wrong attributions and stale statuses are the three failure modes that matter. All three are fixed on report, with no argument about whether the report is worth acting on.