OpenAI reports GPT-5.2 Pro solved an open statistics problem posed at COLT 2019

The case study is the real payload, not the benchmarks

OpenAI leads with two numbers: GPT-5.2 Pro scoring 93.2% on GPQA Diamond and GPT-5.2 Thinking solving 40.3% of FrontierMath (Tier 1–3) problems, which it calls a new state of the art. Both are reported with specific conditions — GPQA with no tools and maximum reasoning effort, FrontierMath with a Python tool enabled and maximum reasoning effort.

But the part of the post that actually shows the model doing something new is the case study on learning-curve monotonicity. That's where the claim moves from 'answered graduate-level questions' to 'contributed a proof to a previously unresolved research problem.'

What the open problem actually was

The problem traces back to an open question posed at the Conference on Learning Theory in 2019 by Viering, Mey, and Loog: if you collect more data, do your results reliably improve? Researchers had shown that even simple setups can have non-monotonic learning curves, where more data increases expected error.

The specific gap GPT-5.2 Pro addressed was the cleanest textbook case: a correctly specified model, Gaussian data, known mean, unknown standard deviation. Small perturbations to that setup were already known to break monotonicity, but the core case was open. The resulting paper, 'On Learning-Curve Monotonicity for Maximum Likelihood Estimators,' shows learning is predictably improved by more data in this setting.

The workflow OpenAI is careful to describe

OpenAI stresses how the proof was obtained, not just that it was obtained. The authors did not hand the model a proof outline or intermediate arguments. They asked GPT-5.2 Pro to solve the open problem directly, then verified the result — including review by external subject-matter experts. Follow-up questions pushed the model to extend the result to higher-dimensional settings and other common statistical models.

Throughout, the human role stayed focused on verification and clear writing, rather than supplying mathematical scaffolding.Montana Labs

That sentence is the load-bearing claim of the whole post. It's a specific division of labor: the model generates the structured argument, humans own correctness and exposition. OpenAI also names the limits plainly — these systems 'are not independent researchers,' can make mistakes, and can rely on unstated assumptions.

The interface this implies: prompt, then verify

For anyone building tools on top of these models, the honest takeaway is about the human-facing surface, not the leaderboard. The described mode of work has two touch points: a researcher poses a well-formed problem, and then a separate, deliberate verification pass by domain experts. The model's output is treated as a detailed argument that 'merit[s] careful human study and refinement,' not an answer to accept.

That shape has consequences for what a good frontend looks like in theoretical domains. The valuable affordances aren't chat convenience; they're structured proof presentation, traceable assumptions, and a review workflow that makes external validation cheap. OpenAI is explicit that reliable progress 'depends on workflows that keep validation, transparency, and collaboration firmly in the loop' — which is a product requirement as much as a research caution.

The specific implication of this announcement: in axiomatic fields like mathematics and theoretical computer science, the bottleneck OpenAI is describing has shifted from generating candidate arguments to verifying them. Teams that want to use GPT-5.2 for this kind of work should invest their engineering effort in the verification interface, because that is the step the post says remained entirely human.