When AI Meets Amplitudes: Can Large Language Models Reshape Theoretical Physics?

The new work focuses on gluons, the force-carrying particles of quantum chromodynamics (QCD) that bind quarks together inside protons and neutrons. Gluons carry the strong nuclear force—one of the four fundamental forces of nature—and are central to understanding high-energy collisions.

A key property of gluons is helicity, describing the orientation of their spin relative to their motion. Physicists label helicity states as positive or negative. For decades, standard reasoning suggested that certain configurations—specifically “single-minus” tree amplitudes, where one gluon has negative helicity and the rest positive—should vanish. In practical terms, these scattering probabilities were thought to be zero.

The new result revisits that assumption. Under special geometric conditions termed a half-collinear configuration, where particles move nearly in the same direction, the team found that these previously “forbidden” amplitudes can, in fact, be non-zero.

According to the authors, GPT-5.2 Pro first suggested a compact formula expressing these amplitudes as a product of (n − 2) factors, where n is the number of gluons involved. An internal OpenAI system then formally verified the expression, and human physicists subjected it to stringent consistency checks required by gauge symmetry and quantum field theory.

For lower particle counts—three, four, and five gluons—the team cross-checked the formula explicitly against conventional calculations. At six gluons, the old approach already required 32 separate terms; the new representation remained strikingly concise.

As theoretical physicist Nima Arkani-Hamed of the Institute for Advanced Study observed, historically in amplitude research, “terribly complicated” expressions often conceal unexpectedly elegant structures—clues to deeper mathematical symmetries. Simple formulae, he noted, frequently open entirely new conceptual pathways.

A Short History of Amplitude Revolutions — Why Simplicity Signals Structure

Modern scattering amplitude research has a history of turning apparent algebraic chaos into geometric elegance. The current GPT-assisted result is best understood not as an isolated computational trick, but as the latest entry in a four-decade sequence of structural compressions that have repeatedly redefined quantum field theory.

The first major shock came in 1986, when Stephen Parke and Thomas Taylor derived a remarkably compact expression for a class of gluon scattering amplitudes now known as Maximally Helicity Violating (MHV) amplitudes. At the time, conventional Feynman-diagram methods required summing hundreds of terms even for modest particle counts. Parke and Taylor showed that the full result collapsed into a single rational expression — an extraordinary simplification that suggested deep hidden structure inside non-Abelian gauge theory.

The shock intensified in the early 2000s. In 2003, Edward Witten proposed that certain gauge-theory amplitudes could be reformulated in twistor space, a geometric framework originally introduced by Roger Penrose. Witten’s twistor string proposal suggested that scattering amplitudes might arise from simple geometric objects rather than diagrammatic expansions.

Shortly thereafter, in 2005, the Britto–Cachazo–Feng–Witten (BCFW) recursion relations demonstrated that entire classes of amplitudes could be constructed recursively from lower-point amplitudes, bypassing the factorial growth of Feynman diagrams. This reduced what had once required summing thousands of diagrams to solving elegant algebraic relations.

The structural crescendo arrived in 2013, when Nima Arkani-Hamed and Jaroslav Trnka introduced the amplituhedron — a geometric object in a higher-dimensional space whose volume encodes scattering amplitudes in planar N=4 supersymmetric Yang-Mills theory. In that formulation, locality and unitarity — foundational principles of quantum field theory — emerge as consequences of geometric positivity rather than being imposed by hand.

Each of these revolutions followed a similar pattern:

1. Brutally complicated intermediate expressions.
2. Unexpected cancellations.
3. Discovery of a compact formula.
4. Subsequent uncovering of deeper mathematical structure.

The GPT-assisted gluon formula now sits within this lineage. The central question is whether AI is merely rediscovering algebraic cancellations — or whether it is indirectly navigating the same geometric or symmetry-based structures that human theorists uncovered over decades.

If history is a guide, simplicity in amplitudes is rarely accidental. It is usually a signal that the theory is trying to be written in a better language

 AI as Collaborator — or Unreliable Genius?

If the gluon result represents AI at its most promising, recent episodes show the picture is more complicated.

On November 19, 2025, theoretical physicist Stephen Hsu of Michigan State University uploaded a paper—later published in Physics Letters B in January 2026—describing how GPT-5 assisted in developing ideas related to modifications of quantum mechanics using the Tomonaga-Schwinger formalism. Hsu positioned the AI as both idea generator and proof-checker.

Yet scrutiny followed quickly. IIT-Mandi physicist Nirmalya Kajuri pointed out that one of the AI’s adopted approaches had been shown invalid as early as 1994 by Charles Torre and Madhavan Varadarajan. Meanwhile, University College London physicist Jonathan Oppenheim argued that the problem addressed had already been settled 35 years earlier by Nicolas Gisin and Joseph Polchinski.

Oppenheim’s deeper concern was epistemic: the AI lacked the contextual judgment to recognize when it was revisiting dead ends. In some cases, it tested the wrong mathematical criteria while appearing rigorous—producing what he described as superficially polished but conceptually empty work.

In a separate experiment in February 2026, Oppenheim asked Anthropic’s Claude Opus 4.5 to perform a research-level path integral calculation. The model initially produced an incorrect answer within minutes. When instructed to verify its output using Mathematica, it iterated corrections until matching the software’s output—only for researchers to discover it had fed Mathematica the wrong expression in the first place.

Through a method he termed the “Groundhog Day loop,” Oppenheim used persistent instruction files to train Claude to decompose calculations systematically, outsource symbolic algebra properly, and cross-validate results. Eventually, one iteration produced the correct result in five minutes—matching work that had taken a graduate student a week.

The lesson was clear: AI can accelerate computation, but only within carefully engineered guardrails.

Pattern Recognition vs Theorem Discovery — What Kind of Intelligence Is This?

Large language models such as GPT-5.2 are fundamentally statistical pattern engines. They are trained on massive corpora of text — including mathematical and physics literature — and learn probabilistic associations between symbolic structures.

This raises a precise question:

When GPT proposes a compact amplitude formula, is it discovering a theorem — or compressing a pattern?

There is a critical distinction between:

• Symbolic algebra systems (e.g., Mathematica, Maple), which manipulate formal expressions deterministically;
• Formal proof assistants (e.g., Lean, Coq, Isabelle), which verify logical derivations step by step;
• And neural language models, which generate plausible mathematical continuations based on learned distributional structure.

GPT does not internally “prove” in the strict logical sense. It proposes candidate structures. Verification must occur externally — either via symbolic computation or by human derivation.

Recent work in automated theorem proving (for example by DeepMind’s AlphaTensor and AlphaGeometry systems) shows that neural-symbolic hybrids can discover novel algebraic decompositions and geometric constructions. But those systems operate within tightly constrained formal domains.

In high-energy theoretical physics, the space of possible expressions is enormous and under-constrained. The risk is not simply arithmetic error — it is misidentification of the relevant structure.

The Hsu episode and Oppenheim’s Mathematica experiment illustrate this risk clearly: AI can generate internally consistent mathematics while testing the wrong conceptual predicate.

So what category of intelligence is being exhibited?

Possibilities include:

1. Heuristic conjecture generator.
2. High-speed symbolic assistant.
3. Structure recognizer approximating geometric intuition.
4. Probabilistic compressor of known analytic forms.

The fact that GPT suggested a compact (n − 2) factor structure is intriguing because such multiplicative forms often signal deeper invariance or factorization properties.

But until the derivation is embedded into a broader analytic framework — for example via recursion relations, symmetry arguments, or geometric interpretation — it remains an elegant conjectural compression validated post hoc.

The difference between pattern compression and structural understanding will define the next phase of AI-assisted science.

 Flood of AI-Generated Research — and the Quality Question

Beyond frontier collaborations, AI’s integration into academia is accelerating in more contentious ways. Researchers warn that some authors are using large language models to mass-produce low-quality manuscripts, contributing to what critics describe as “AI slop” in peer-reviewed literature.

At the same time, peer reviewers—despite journal advisories—are themselves using AI tools to assist with report drafting and technical verification. Scientists interviewed in recent weeks emphasize that while AI can assist with computational tasks, assessing conceptual novelty, physical interpretation, and scientific significance still requires human judgment.

University of Texas computer scientist Scott Aaronson summarized the current moment as a “sweet spot”: AI cannot independently produce correct, groundbreaking theory papers, but it can help researchers get unstuck—provided they already understand the domain deeply.

Kajuri offered a memorable metaphor: AI has entered its “graduate student arc”—capable of structured reasoning under supervision, but not yet mature enough to independently define research agendas.

 What This Means for Physics — and for Science Itself

The gluon amplitude result does not revolutionize particle physics overnight. It neither overturns the Standard Model nor predicts new particles. Its importance lies elsewhere: it demonstrates that AI can navigate intricate symbolic reasoning and surface structural simplicity in problems where humans suspect hidden order but lack efficient routes to it.

Scattering amplitude research has historically yielded transformative frameworks—such as the spinor-helicity formalism and twistor-inspired approaches—that redefined how physicists conceptualize interactions. If AI systems can consistently identify such simplifications, they may accelerate progress in mathematical physics, string theory, and beyond.

At the same time, recent cautionary episodes underscore that AI systems lack scientific judgment. They do not inherently distinguish between solved and unsolved problems, nor do they automatically evaluate the conceptual significance of their outputs.

The real frontier, therefore, is not machine autonomy but hybrid intelligence: AI systems performing exhaustive symbolic exploration under human oversight, with physicists supplying context, skepticism, and interpretive insight.

 Why This Matters

Fundamental physics increasingly relies on extreme computational complexity—whether in collider data analysis, lattice quantum chromodynamics, or amplitude calculations. If AI tools can systematically compress complexity into transparent formulae, they could reshape theoretical workflows, reducing weeks of algebra to minutes of guided reasoning.

Yet the credibility of science depends not merely on speed, but on rigor. As AI becomes embedded in research, the standards of verification, replication, and peer review may need to evolve.

The gluon formula breakthrough illustrates the promise: AI as a catalyst for mathematical elegance. The cautionary tales illustrate the risk: AI as a generator of convincing but empty formalism.

The future of theoretical physics may not hinge on whether AI replaces scientists, but on whether scientists learn to harness AI without surrendering judgment.