GPT-5.2 Pro just solved a quantum gravity puzzle

For decades, physics textbooks took a specific rule for granted: at the simplest tree level, certain graviton interactions—specifically single-minus amplitudes, where one graviton has negative helicity and the rest are positive—must vanish.

Turns out, that textbook assumption was incomplete. In a new preprint co-authored by researchers from Harvard, Cambridge, the Institute for Advanced Study, Vanderbilt, and OpenAI, the team revealed that these single-minus graviton tree amplitudes are actually nonzero under a special momentum alignment called the half-collinear regime.

Even more striking: GPT-5.2 Pro was central to cracking the math.

The half-collinear regime

Scattering amplitudes calculate the probabilities of particle interactions without forcing physicists to track every intermediate step through endless Feynman diagrams. The old belief that single-minus interactions vanish assumed generic particle motion.

Under the half-collinear regime, however, standard cancellations break down. Instead of vanishing, the amplitudes survive as well-defined mathematical distributions. The authors proved these formulas using symmetry principles and recursion relations, eventually linking the behavior directly to an infinite-dimensional symmetry known as w-(1+∞), originally introduced by Roger Penrose fifty years ago.

Connecting this symmetry to gravitons marks a tangible step toward bridging quantum mechanics and Einstein's general relativity.

An unexpected graph theory trick

The real breakthrough in methodology came when translating an earlier gluon finding over to gravitons. While gauge theory and gravity share conceptual roots, their actual calculations are notoriously distinct.

The researchers fed the completed gluon paper into GPT-5.2 Pro as context, asking it to derive the corresponding quantum gravity amplitudes—a task that would normally take theoretical physicists significant time. The model surprised everyone by solving the problem using the directed matrix-tree theorem from graph theory, while simultaneously producing a remarkably coherent first draft of the paper. Human theorists then spent the bulk of their time verifying the analytic proofs and running consistency checks against known physical limits.

Why it matters

This is one of the cleanest real-world demonstrations of AI operating as a genuine scientific co-pilot, rather than just a summarizer or proofreader.

What makes this project notable is the shift in human workload. As the authors highlighted, the bottleneck flipped overnight: formulating conjectures and testing mathematical bridges became nearly instant, leaving researchers to spend almost all their time on validation and analytical verification.

We think OpenAI's strategy here is clever because it anchors the model to a concrete, verifiable mathematical domain rather than vague discovery. The risk, of course, is verification debt. As frontier models get faster at cross-pollinating complex formulas across distinct branches of math and physics, human expertise may soon become the hard bottleneck in vetting the tsunami of non-obvious proofs AI tosses back over the fence.