A team of physicists used ChatGPT to help crack a long-standing problem in quantum field theory, producing a new closed-form expression for single-minus gluon tree amplitudes that specialists had assumed were zero. The result, detailed in a February 2026 preprint, is one of several recent cases in which large language models have moved from novelty to working tool inside particle physics, assisting with everything from theoretical conjectures to the hands-on tuning of particle accelerators.
An AI Chatbot Joins a Gluon Amplitude Proof
For decades, single-minus tree-level n-gluon amplitudes were treated as vanishing quantities in standard calculations. That assumption held under ordinary kinematic conditions, but a group of researchers showed that these amplitudes can in fact be nonzero in a specific complexified regime. The team used an LLM-assisted workflow to guide their exploration, with the chatbot suggesting algebraic paths and candidate expressions that the physicists then verified by hand. The preprint describes how the AI functioned less as an oracle and more as a creative sparring partner, proposing conjectures that human collaborators refined into a rigorous result.
The finding matters because gluon scattering amplitudes are central to predictions at colliders like the Large Hadron Collider. Getting them right, or discovering that accepted shortcuts were hiding real structure, has direct consequences for how theorists model particle interactions and interpret precision data. A narrative account syndicated from the Harvard Gazette added timeline details and plain-language context about why the amplitude result drew attention in its niche community, while also highlighting the awkwardness of assigning intellectual credit to a chatbot. For now, the authors list remains human, but the episode has intensified debate over whether future conventions should recognize AI tools more formally when they contribute to derivations that might not have emerged otherwise.
Tuning Accelerators With Natural Language
Theory papers are not the only place where LLMs have gained traction in high-energy physics. A peer-reviewed study in Science Advances demonstrated that natural-language prompting can drive real optimization of particle accelerator settings, turning a chat interface into a control-layer assistant. The researchers structured prompts so that an LLM could interpret beam diagnostics, propose adjusted machine parameters, and iterate toward better performance. Their technical preprint on the accelerator workflow details the prompt designs, explains how text responses were parsed into concrete machine settings, and documents the baselines used for comparison against traditional tuning procedures.
In controlled tests, the system outperformed manual tuning, reaching comparable or better beam quality with fewer iterations and less operator intervention. A Nature research highlight independently characterized this work, describing LLMs as functioning like co-pilots in accelerator control rather than autonomous drivers. That framing is apt: operators still make final decisions and can veto unsafe suggestions, but the AI handles the tedious translation between human intent and machine configuration. For facilities that run around the clock and rely on small teams of operators, that kind of assistance could meaningfully reduce downtime. The broader implication is that LLM-style chat systems are now practical tools not just for writing or coding but for critical particle-physics infrastructure itself, provided they are embedded in carefully engineered safety envelopes.
Transformers Generate Theories and Tag Jets
Beyond accelerator operations, transformer architectures are being applied to two of particle physics’ core intellectual tasks: building candidate theories and classifying experimental data. On the theory side, one research group trained a transformer to map particle content to candidate Lagrangians while respecting gauge-symmetry constraints, effectively automating a step that theorists normally perform by hand over weeks or months. In their study, the authors report quantitative benchmarks for theory generation and release open model artifacts, showing that structured idea generation in theoretical physics is no longer purely a human activity. The model does not replace conceptual insight, but it can rapidly enumerate consistent possibilities, helping researchers survey large spaces of models that would be impractical to explore manually.
On the experimental side, the Particle Transformer architecture, known as ParT, was introduced alongside the large public JetClass dataset for jet tagging. This dataset includes detailed simulation provenance and covers multiple jet classes, giving the community a shared benchmark for comparing machine-learning approaches to collider data. The original JetClass and ParT study showed that transformer-based networks can exploit the full set of low-level particle features in a jet, delivering state-of-the-art performance on classification tasks that are central to many LHC analyses. Building on that foundation, newer efforts such as HEP-JEPA adapt joint-embedding predictive architectures with self-supervised pretraining for collider-physics workflows, as described in a dedicated foundation-model preprint on high-energy data. These models are not general chatbots; they are purpose-built systems tuned to the statistical patterns and symmetries that characterize collision events.
Domain-Tuned Models Challenge Commercial AI
A recurring question in this space is whether general-purpose commercial models like ChatGPT are good enough for specialized physics tasks, or whether the field needs its own fine-tuned alternatives. Researchers behind a project called FeynTune addressed this directly by fine-tuning large language models on high-energy theory corpora and then benchmarking domain-tuned LLMs against commercial systems. On tasks such as deriving Feynman rules, simplifying amplitudes, and answering technical questions about gauge theories, the customized models matched or exceeded the performance of closed-source chatbots, despite having fewer parameters and being trained on more focused datasets.
This matters because commercial AI tools carry costs, usage limits, and data-privacy concerns that make them awkward fits for sustained research programs, especially when proprietary models cannot be inspected or reproduced. A fine-tuned open model trained on the right corpus can be deployed locally, iterated on freely, and shared across collaborations without licensing friction, aligning better with the open-science norms of high-energy physics. The tradeoff is that building and maintaining such models requires compute resources, data curation, and ongoing evaluation that not every research group can afford. For now, the practical reality is a hybrid: physicists reach for ChatGPT or similar systems when they need quick exploratory help or broad-domain knowledge, and turn to specialized models like FeynTune or HEP-JEPA when precision, interpretability, and reproducibility are non-negotiable.
What AI Still Cannot Do for Physics
For all the progress, a clear limitation runs through every one of these applications: AI outputs require human verification at every step. In the gluon-amplitude work, the chatbot could suggest algebraic manipulations, but only expert scrutiny could confirm that the resulting expression satisfied unitarity, gauge invariance, and known limits. In accelerator control, safety interlocks and human oversight are mandatory, because even a small misconfiguration could damage equipment or compromise experimental runs. And in theory generation, a model that proposes Lagrangians is only as useful as the physicists who can interpret which ones are physically meaningful, consistent with data, and worth developing into full-fledged scenarios.
These constraints point to a division of labor rather than an impending handover of physics to machines. Large language models excel at pattern recognition, rapid search over symbolic possibilities, and translation between natural language and formal representations, but they lack the grounding in experiment, the long-term intuition about which questions matter, and the responsibility for consequences that define scientific judgment. As tools like ChatGPT, ParT, HEP-JEPA, and FeynTune become more capable, the challenge for the field will be to integrate them in ways that amplify human insight without eroding standards of rigor. The recent successes in amplitudes, accelerator tuning, theory generation, and jet tagging suggest that such a partnership is not only possible but already underway, provided physicists remain clear-eyed about what AI can and cannot do.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
