A previously dismissed class of particle interactions may exist, thanks to a breakthrough facilitated by artificial intelligence. Researchers from the Institute for Advanced Study, OpenAI, Vanderbilt University, Cambridge University, and Harvard University have announced a finding that challenges long-held assumptions about gluon interactions, the forces that bind quarks within protons and neutrons.
The discovery, detailed in a pre-print study posted to arXiv, centers on “single-minus” configurations of gluons – arrangements where one gluon has a negative helicity while the others have positive helicity. Standard calculations had suggested these interactions would effectively vanish at the most basic level of analysis, known as “tree level.” However, the research team demonstrated that under specific conditions, these interactions not only exist but can be described with a surprisingly compact mathematical formula.
The key to this breakthrough was the application of OpenAI’s GPT-5.2 Pro. After researchers manually calculated amplitudes for small numbers of gluons, resulting in complex expressions derived from Feynman diagrams, the AI was used to simplify those expressions. From these simplified cases, GPT-5.2 Pro identified a pattern and conjectured a general formula applicable to any number of gluons. This conjecture was then formally proven and independently verified by the human researchers.
The significance lies not just in the physics itself, but in the methodology. Traditionally, identifying such patterns required years of experience and deep intuition. The AI’s ability to infer a general formula from specific cases represents a potential shift in how theoretical physics research is conducted. As Nima Arkani-Hamed, Professor of Physics at the Institute for Advanced Study, noted in a statement accompanying the release of the study, finding a simple formula has historically been a “fiddly” process, and one that “might be automatable by computers.”
The study focuses on scattering amplitudes, mathematical objects that encode the probability of particle collisions. As the number of particles involved increases, the complexity of calculating these amplitudes grows exponentially, making even seemingly simple interactions computationally intractable. The AI-assisted approach offers a way to bypass this complexity, revealing underlying simplicity that had previously remained hidden.
Specifically, the researchers found that when gluon momenta are aligned in a particular way – a “half-collinear regime” – the standard arguments for the vanishing of these interactions break down. This alignment leads to a coordinated vanishing of certain mathematical quantities, allowing the interaction to persist. The resulting formula provides a systematic way to build up complex multi-particle interactions from simpler ones, using a method known as the Berends-Giele recursion relation.
The simplification is particularly dramatic in a specific arrangement known as region R1, where a single negatively spinning gluon transforms into many positively spinning ones. In this scenario, the complex web of interaction diagrams collapses into a concise formula based on simple plus, minus, or zero values, depending on particle alignment.
While the current results apply to tree-level amplitudes and specific kinematic regimes, the implications extend beyond gluons. The researchers suggest the approach could be generalized to gravitons – the hypothetical carriers of gravity – and supersymmetric extensions of the Standard Model. However, they acknowledge that loop corrections, which account for quantum fluctuations, remain significantly more complex.
The methodological implications are perhaps the most far-reaching. The researchers propose a model where AI generates conjectures, which are then rigorously proven and validated by human scientists using established analytical methods. This collaborative approach could unlock new avenues of discovery in areas of theoretical physics characterized by hidden simplicity within complex algebra.
Nathaniel Craig, Professor of Physics at the University of California, Santa Barbara, described the study as “a glimpse into the future of AI-assisted science,” emphasizing the potential for physicists and AI to work together to generate and validate new insights. He added that the work “provides a template for validating LLM-driven insights and satisfies what we expect from rigorous scientific inquiry.”
The research team included Alfredo Guevara, of the Institute for Advanced Study; Alexandru Lupsasca of Vanderbilt University and OpenAI; David Skinner of the University of Cambridge, Andrew Strominger of Harvard University and Kevin Weil, of OpenAI. The study is a significant step towards integrating AI into the core processes of scientific discovery, potentially accelerating progress in fundamental physics.
the study is currently available as a pre-print on arXiv, meaning it has not yet undergone the full peer-review process. Peer review is a crucial step in the scientific process to ensure the validity and reliability of research findings.
