Researchers have developed a research agent based on GPT-5.2 Pro, designed to solve challenging problems in experimental mathematics, with plans to extend the same framework to computational physics in the future.

In its first deployment, the agent achieved a new best-known result on the spherical coding problem (parameters n=11, N=432), a result that has been verified against the benchmark library maintained by Henry Cohn (MIT).

The agent's strategy successfully broke through a numerically 'stuck' configuration that had previously resisted optimization attempts, ultimately optimizing the cosine value to approximately 0.49422771.

Notably, the agent achieved this improvement in just about an hour of autonomous exploration, whereas discovering and optimizing similar configurations previously might have required significant human effort and large-scale computing resources.

Verification results can be found at: [Spherical Codes Library](https://spherical-codes.org/)

Some netizens pointed out that although this optimization targeted only a specific numerical value, the AI's ability to customize algorithms to solve problems without general solutions represents an important advancement.

However, some users mentioned that GPT-5.2 Pro has made errors when processing HTML files. The developer responded that the AI lacks sensory verification capabilities and can only self-correct by writing and running tests. If AI gains more perceptual abilities in the physical world in the future, its potential will be much greater.

Regarding the significance of the optimization results, some netizens asked whether the tiny difference of 0.00000004 is important. The developer explained that the difference itself is not significant, but the key point is that the AI could still find new directions for improvement when approaching the optimal solution, indicating that its exploration strategy differs from and is effective compared to human approaches. Such high-search-space problems are widely used in industry, and this breakthrough by AI预示着 its potential in fields like material optimization.

Although GPT-5.2 Pro still makes mistakes on certain tasks (such as LaTeX format generation), in another run, it proactively took screenshots to check page formatting, showing a certain degree of foresight.

Ultimately, this achievement not only demonstrates AI's capabilities on complex mathematical problems but also provides new directions for the development of autonomous research agents.