GPT-5 Mathematical Discovery

GPT-5 Accelerates Solution to Longstanding Optimization Problem

OpenAI's GPT-5 enabled Professor Ernest Ryu of UCLA to solve a 40-year-old open problem regarding the stability of the Nesterov Accelerated Gradient (NAG) method. By acting as a creative collaborator, GPT-5 accelerated the exploration of potential proofs, condensing weeks of potential manual work into approximately twelve hours of collaboration over three days.

The Nesterov Accelerated Gradient (NAG) Problem

The core mathematical challenge involved understanding why the Nesterov Accelerated Gradient (NAG), introduced by Yurii Nesterov in 1983, remains stable despite providing a significant speed boost to algorithm convergence.

Technical Context of NAG

  • Function: NAG is an optimization method that uses a "look-ahead" prediction to calculate gradients at a future point rather than the current state.
  • Benefit: This approach allows algorithms to converge faster and manage oscillations more effectively as they approach a function's minimum.
  • The Mystery: For decades, researchers observed that the added momentum from NAG did not introduce instability, but they lacked a formal mathematical proof to explain why this behavior occurred.

GPT-5 as a Mathematical Collaborator

Professor Ryu utilized GPT-5 not as an autonomous solver, but as an exploratory tool to surface ideas and techniques from across a vast range of mathematical literature, including adjacent subfields.

The Collaborative Workflow

  • Idea Generation: GPT-5 proposed unconventional approaches and restructured equations, some of which were "out of the blue."
  • Rapid Iteration: Ryu used the model to quickly propose and discard variations of an idea, which is typically a cognitively exhausting process for humans.
  • Verification and Filtering: Ryu acted as the primary validator, identifying promising directions and discarding flawed reasoning. He noted that GPT-5 often produced arguments that looked plausible but were incorrect upon inspection.
  • Error Mitigation: To minimize accumulated errors, Ryu started new chat sessions when asking the model to check work, rather than continuing a single long conversation.

The Turning Point

The solution emerged when GPT-5 suggested a specific way of restructuring the equations governing the NAG method. While the suggestion was not correct as written, Ryu recognized a meaningful structural feature within the suggestion. He developed this feature rigorously on his own and used targeted questions to test the proof's viability, which became the backbone of the final proof.

Implications for Mathematical Research

This case study demonstrates that the partnership between AI and subject-matter experts can significantly accelerate discovery by expanding the theoretical tools available to a researcher.

Human-AI Synergy

  • Expertise Requirement: The process required Ryu's deep expertise in optimization theory to discern valuable insights from the model's errors.
  • Psychological Impact: The rapid flow of ideas provided by GPT-5 prevented the typical frustration that leads researchers to abandon difficult problems, keeping the problem feeling "within reach."
  • Academic Integration: Ryu's resulting pre-print paper (available at arxiv.org/abs/2510.23513) follows traditional scholarly styles to serve as a case study for classical mathematicians, though GPT-5 is credited in the title and abstract.

"The way math research works is you have ideas, and whenever youหรือ colleagues come up with a rough idea, you have a sense of whether or not it is going to work. This is where the partnership between artificial intelligence and humans can work especially well."

Conclusion

GPT-5 did not invent new mathematical principles but excelled at wielding existing tools and connecting disparate ideas from a massive scale of reading. The success of this discovery relied on a human expert's ability to verify reasoning and guide the exploration process.

Sources