Case Studies

OpenAI Quietly Pulls Three Math Papers

Three arXiv papers vanished from a public OpenAI repo, and Hacker News noticed. Whether this signals shifting research priorities or a strategic pullback from open-source commitments, the timing lines up too neatly with OpenAI's closed-source reasoning models.

Font size

A Hacker News thread blew up last night over OpenAI pulling three math papers: 308 upvotes, 295 comments. Paper withdrawals aren't unusual. What is unusual is one coming from OpenAI, on a public GitHub repo, with every step traceable through git log.

What Actually Got Pulled

The trail starts at the history.md file in OpenAI's public openai/math repository. The file was originally used to track research progress around the PRM800K dataset, but several commits from late 2025 show three papers quietly marked as withdrawn.

The rough timeline:

  • Let's Verify Step by Step (2023), the core methodology paper behind PRM800K
  • LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks (early 2024), work on using LLMs as planning aids
  • weak-to-strong generalization, a math-direction extension (mid 2024) that followed up on the well-known GPT-4-era paper

I couldn't find the full withdrawal statements in the repo (they may only live on arXiv), but the history.md records are unambiguous: WITHDRAWN. The corresponding timestamps are visible in the GitHub commit history, with changes rolled out sometime between November and December 2025.

Was this a genuine withdrawal over problems, or simply a delisting from a public repo? The latter seems more likely. The corresponding arXiv pages I could locate are still live; OpenAI's own repo just no longer links to them.

Why This Hit the Front Page

Across 295 comments, the discussion split into two camps.

One camp questioned the work itself. Some users claimed the results in Let's Verify Step by Step don't reproduce, and that the math-task conclusions from weak-to-strong don't hold up. This wasn't the majority view, but a few people posted benchmark numbers.

The other camp pressed on OpenAI's open-source commitments. One high-upvoted reply put it bluntly:

"OpenAI's open-source stance has always been 'we open-source what we want to open-source.' This is them formally retracting what they don't want to open-source."

Another comment dug into the deeper pattern:

"Three math papers pulled, three reasoning model (o1/o3) iterations. Math reasoning capability now lives in the products, not the papers."

That comment pulled 200+ upvotes and captured the room's consensus. The withdrawals line up perfectly with OpenAI's push to productize reasoning behind closed doors.

After PRM800K, OpenAI's Math Strategy Shifted

Looking back at OpenAI's key moments in math reasoning, the recent pulls aren't a one-off.

2022–2023: The PRM800K era. "Let's Verify Step by Step" proposed using a Process Reward Model (PRM) to score each reasoning step. That paper directly produced the PRM800K dataset (800K step-level annotations). OpenAI was generous with the academic community during this period: the dataset was public, model weights were semi-public, and papers were detailed. The methodology got widely cited, with follow-on work from NVIDIA's math evals and several DeepMind efforts building on the same ideas.

Mid-2024: o1 arrives. The reinforcement-learning version of chain-of-thought became a product with the release of o1-preview. A paper? Almost none. The technical report was eight pages, basically a product brochure.

Late 2024–2025: o3, o3-mini iterations. Closed-source models kept getting stronger, topping ARC-AGI and math olympiad benchmarks. The pace of public papers clearly slowed. Even methodological breadcrumbs got thinner.

Late 2025: Three math papers withdrawn. This isn't "we stopped caring about math." It's "math reasoning capability now lives in the product, not the paper."

The timeline is too neat to be coincidence.

Paper Withdrawals Aren't Rare. OpenAI Paper Withdrawals Are.

Quick context: it's genuinely rare for a top AI lab to actively delist papers from a public repo.

  • DeepMind occasionally withdraws papers, but usually for small arXiv-side fixes. Public repos (like alphafold2) are rarely touched.
  • Anthropic's public repos are mostly engineering tools (constitutional AI, prompt utilities), with few academic papers.
  • Meta pulling a Llama paper? Essentially unheard of. The Llama technical reports stay up; flagged issues get revisions, not withdrawals.

So what's different here?

The repo is public. openai/math is a GitHub public repository, and every action leaves a commit log. Every file modification, who made it, and the commit message are visible to anyone.

In other words: when other labs withdraw papers, you can only guess. When OpenAI withdraws papers, you can look up the timestamp and reason for every change.

This transparency cuts both ways: any move gets amplified under the spotlight. An HN user posted a git log diff and walked through the citation graph before and after the withdrawal. That kind of analysis is impossible at any other lab.

So the virality here isn't about "a withdrawal," it is about "OpenAI withdrawing papers from a public repo." The latter is rarer, and far more interpretable.

My Read

The real signal isn't the withdrawal itself. It's that OpenAI is reclassifying research assets from "public goods" into "product moats."

A few things point to this:

  1. Math reasoning is now the core selling point of the o-series models. Letting PRM800K-style methodology continue evolving in public would hand competitors the training playbook for the next o-model.
  2. Public paper output and closed-source product iteration move in opposite directions. High paper output in 2023, a pivot toward products starting in 2024, and withdrawals of earlier work by late 2025, which is the motion of a team clearing out its "public goods" inventory.
  3. The repo's transparency makes the moves more visible. If this had happened inside a closed-organization company, nobody would have known.

Things I'm less sure about:

  • OpenAI hasn't stated the real reason for the withdrawals. It could be legal, it could be that the technical conclusions didn't hold up, or it could simply be a strategic adjustment.
  • The impact on academia is hard to estimate. PRM800K is still on HuggingFace, citations won't disappear, but the expectation that "the OpenAI math group is still working on these directions" is gone.
  • Will this kind of withdrawal spread to other OpenAI public repos? openai/evals, openai/gpt-2, and others are untouched for now, but the precedent is there.

On the open-source side, my guess is that a few independent teams will fork PRM800K's methodology and keep building in the short term. Over the longer horizon, open-source alternatives for basic math reasoning (Qwen-Math, DeepSeek-Math) will draw more attention, because everyone's now aware that the major labs won't keep investing public resources in this direction.

Final thought: pulling three papers isn't a big deal in isolation. But the timing and the method just happen to expose how OpenAI is redrawing the line between "what belongs to the product" and "what belongs to the community." Where that line ends up matters far more than the papers themselves.


References:

Source: https://github.com/openai/math/blob/main/history.md