AI helped compute graviton amplitudes in quantum gravity

Theoretical physics has long been considered a domain exclusively for the human mind — a field where intuition and years of specialized training are indispensable. A new preprint published by a group of researchers challenges this axiom: the large language model GPT-5.2 Pro has for the first time helped derive and verify non-zero tree-level graviton amplitudes within the framework of quantum gravity. This is not merely a technical curiosity — it is a signal that artificial intelligence is beginning to truly work where previously only a narrow circle of specialists could act.

To grasp the significance of the result, one must delve into the context. Scattering amplitudes are mathematical objects that describe the probabilities of interactions between elementary particles. In quantum field theory, their computation using classical Feynman diagrams has long been known as an extremely labor-intensive process: the number of terms grows factorially with the number of participating particles.

The method of so-called single-minus amplitudes — a special class of helicity configurations where exactly one of the external momenta has negative helicity — allows one to significantly reduce analytical work. However, until recently, this method was applied predominantly to gluons in gauge theories, such as quantum chromodynamics. Extending the approach to gravitons — particles that mediate gravitational interaction within quantum gravity — presented a separate non-trivial task.

Gravitons in the theoretical sense are considerably more complex than gluons. Their scattering amplitudes, even at the level of tree diagrams without accounting for quantum corrections, generate expressions of truly formidable complexity. It was here that GPT-5.

2 Pro entered the scene. The researchers employed the model not as a replacement for a human physicist, but as an analytical assistant capable of operating with symbolic computations, tracking long algebraic chains, and verifying intermediate results for correctness. The model helped derive specific non-zero amplitudes and verified them by comparing them against known relations — in particular, the Kawabata-Luforó-Stevens relations that connect gravitational and gauge amplitudes.

It is important to emphasize: we are not discussing numerical simulation or parameter tuning — GPT-5.2 Pro worked precisely with symbolic algebra, that very algebra which traditionally requires the involvement of a human with professional qualification.

This distinction is fundamental. In recent years, neural networks have been actively applied in physics to solve differential equations, optimize molecular structures, or accelerate simulations. But all of this involves tasks where the model operates on numbers or approximates functions. Analytical computations in theoretical physics are a fundamentally different matter: here one must manipulate symbolic expressions, understand the structure of Lie algebras, the symmetry group of the problem, and the physical constraints imposed on the result. The fact that a language model succeeds at such work at a level sufficient for inclusion in a scientific preprint represents a qualitative shift in the capabilities of the technology.

For industry and the academic community, this result opens several important perspectives. First, it legitimizes the use of LLMs in fundamental theoretical science — an area where such tools have previously been viewed with skepticism. Second, it points to a specific niche where language models provide real advantage: not replacing the scientist, but accelerating the most routine, though technically complex, stages of work. Cumbersome algebraic manipulations, on which a theoretical physicist might spend weeks, potentially compress to hours. Third, the publication stimulates discussion about verification: how should the community treat results partially obtained with the help of AI? The authors, it would seem, made a bet on transparency — explicitly indicating the role of GPT-5.2 Pro in the work.

Quantum gravity remains one of the principal unsolved problems of fundamental physics. The theory still lacks experimental confirmation, and the analytical apparatus necessary for its development is extraordinarily complex. If language models are capable of taking on part of this analytical burden, the pace of theoretical research may accelerate noticeably. Not because AI understands physics in a human sense — but because it is able to operate with extraordinary precision and speed on those formal structures through which physics expresses itself. This is a preprint worth following: if the results pass peer review, we may be witnessing the beginning of a new working practice in theoretical physics.