OpenAI has released a second round of AI-produced research in mathematics, computer science and physics, weeks after a first batch of manuscripts. It is trying three ways of handling the explaining: maths results go out for researchers to explain, computer science results are public while an outside expert group writes its own account, and the physics release is held for its expert group. Two new manuscripts are on GitHub.
“AI systems are able to give a lot of output that is mathematical in some ways, but is not conforming to important academic standards”
Melanie Matchett Wood, Harvard mathematics professor and advisory group member, as quoted by The Hill from the Harvard Crimson · thehill.com ↗
One claims to establish Sogge's local smoothing conjecture for waves in every spatial dimension. The other claims the Exponential Time Hypothesis is false. That hypothesis is widely believed true and underpins hundreds of papers, so if it fails, better algorithms might exist. Its original proposers are in the group asked to examine the result. The physics claim concerns interacting quantum field theories in four dimensions.
OpenAI says it wants no academic credit. Each group's people get credit for their write-ups and must ensure the AI results acknowledge the researchers and ideas the models drew on.

The Hill's lead picture for its report on OpenAI's earlier batch of math results (7 October)
Image: The Hill
The earlier batch, 372 problems by Engadget's count and 377 results by The Hill's, drew backlash from mathematicians over originality and plagiarism. On 7 October OpenAI formed an independent advisory group, which includes Harvard's Melanie Matchett Wood. Greg Brockman framed that release as speeding scientific discovery.
About OpenAI
This is not the first time an OpenAI maths claim has been challenged. In October 2025 Kevin Weil, an OpenAI manager, posted that GPT-5 had found solutions to ten previously unsolved Erdős problems and made progress on eleven more. Thomas Bloom, who runs the Erdős problems website, called the announcement a dramatic misrepresentation, saying the model had only located published solutions he had not known about.
Weil deleted the post after the criticism and admitted he had misunderstood a researcher's message. Google DeepMind chief Demis Hassabis called the episode embarrassing.
11
Works credited
35
Models indexed
Models on AIPopDB
Next, the physics group is expected to publish, and mathematicians will judge whether OpenAI followed the advice.
Sources · 2

Comments