Geometric algebra as the input language of collider foundation models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Geometric algebra as the input language of collider foundation models".
Title and authors: Kai: So, to summarize what they did with this "Geometric algebra as the input language of collider foundation models," they’re essentially saying that a hard hadron-collider event should be treated as one single geometric object, a graded multivector E <ref:2605.15910#pg2>.
Mira: They claim this representation is the natural mathematical framework because all the observables we use in collider analyses are already organized by its grade decomposition <ref:2605.15910#pg3>.
Lev: The core idea is that a reconstructed event is defined as this multivector E, which has one slot per reconstructed object with a flavor slot too <ref:2605.15910#pg3>.
Kai: They then map the high-level invariants we look for, like the ones mentioned in references four and five, directly to the grade zero sector of this multivector E <ref:2605.15910#pg3>.
Mira: But they also state that the raw covariant coefficients of the higher grades—like those that encode things such as lab-frame or helicity-frame angular variables—are already natural inputs for networks designed to be Lorentz equivariant <ref:2605.15910#pg3>.
Lev: It’s about using these higher-grade components as the input channels for models like L-GATr, which is what they are connecting this framework to in their work on foundation models <ref:2605.15910#pg3>.
Kai: So the main point is that by structuring the event this way, you get a unified input layer for foundation models, unifying what used to be three separate approaches—kinematic data, symmetry information, and object-type labels <ref:2605.15910#pg1>.
Mira: It’s about making sure the model respects Lorentz symmetry without needing to dictate the features itself <ref:2605.15910#pg3>.
Lev: The practical challenge they highlight is that this framework is a common notational tool, not actually new physical content; it’s just a way to organize what we already measure <ref:2605.15910#pg3>.
The paper's summary: Kai: The paper points out some specific improvements they suggest for this approach, mainly around how you use the grade structure for specific tasks <ref:2605.15910#pg2>.
Mira: One big suggestion is that you can design explicit per-grade heads to fine-tune classification and regression tasks; for example, a CP asymmetry test could look directly at the grade four coefficient which they call the "one genuine channel" <ref:2605.15910#pg2>.
Lev: That makes sense from an error correction view because if you know exactly which grade corresponds to which physical quantity, your overhead for extracting that information is much lower <ref:2605.15910#pg4>.
Kai: They also suggest using an input ablation study to figure out exactly which algebraic content actually earns its place in the model input <ref:2605.15910#pg4>.
Mira: For example, they can quantify that a fixed reference blade encoding the beam plane carries structure that the raw momenta don't contain by comparing performance against different input representations <ref:2605.15910#pg4>.
Lev: That helps define the necessary complexity; if you find that certain structures are redundant or not contributing to the observable, you can prune them without breaking things <ref:2605.15910#pg4>.
Kai: And finally, they mention a way to implement a flavor-flavour universality check by using soft weight-tying on feature channels as an S3 permutation among e, mu, and tau tokens of the same lepton type <ref:2605.15910#pg4>.
The paper's improvements: Kai: So to wrap up this paper "Geometric algebra as the input language of collider foundation models," they’re emphasizing that it’s about treating a hard hadron-collider event as one single graded multivector E <ref:2605.15910#pg2>.
Mira: The main implication is that this representation provides a uniform input layer for foundation models, unifying kinematic data and symmetry information into one package <ref:2605.15910#pg1>.
Lev: From the hardware side, it means we have a structured method to incorporate kinematic and symmetry information directly into neural network inputs, which is vital when you’re building something like L-GATr <ref:2605.15910#pg3>.
Kai: They conclude that while the construction is a common notational framework, the key finding is that the only genuine non-trivial invariant content that persists across all these grades is the CP-odd sign of grade four <ref:2605.15910#pg2>.
Mira: They also stress that there are no new Lorentz-invariant scalars unlocked beyond those already present in the initial set, which they bound with what they call the Cayley–Menger collapse <ref:2605.15910#pg3>.
Lev: It’s important to remember their limitation: they state that a reconstructed event is defined as this multivector E, defined in Eq. (four) with one Cl(one three) tensor Vflav slot per reconstructed object <ref:2605.15910#pg3>.
Kai: That means the model respects Lorentz symmetry without needing to dictate the features itself, relying entirely on this geometric structure <ref:2605.15910#pg3>.
Mira: It’s a strong framework for incorporating physical meaning directly into the input specification for foundation models <ref:2605.15910#pg3>.
Lev: We'll keep watching how this helps guide the actual design of those equivariant networks <ref:2605.15910#pg3>.
Conclusion: Kai: So we just talked about how they took a hard hadron collider event and turned it into one single geometric object called a graded multivector E <ref:2605.15910#pg2>.
Mira: Yeah, essentially treating everything—the momenta, the charges, the decay information—as different grades of this one structure <ref:2605.15910#pg3>.
Lev: And they used this multivector E as a uniform input layer for foundation models so you don't need separate machinery for every single layer <ref:2605.15910#pg3>.
Kai: The core idea is that the grade decomposition of this event multivector gives you a fixed interface for every downstream task, which means pre-training on E preserves that grade structure end to end <ref:2605.15910#pg3>.
Mira: That’s important because it lets you have tasks requiring strict invariance reading the grade zero coefficients, while others needing full covariance can just read the whole multivector <ref:2605.15910#pg3>.
Lev: And they showed that the only genuine non-trivial invariant content that persists is the CP odd sign of grade four <ref:2605.15910#pg2>.
Kai: They bound this construction with what they call the Cayley–Menger collapse, which basically says higher grades built from measurements don't add new information beyond what the initial structure already encoded <ref:2605.15910#pg3>.
Mira: But they also pointed out that a grade four object encoding structure that the raw momenta don't contain, like how you single out the laboratory frame by way of the beam direction, that structure is preserved <ref:2605.15910#pg3>.
Lev: So what they’re really showing us is a structured method for incorporating kinematic and symmetry information directly into neural network inputs <ref:2605.15910#pg3>.
Kai: The implication is that this framework unifies three previously separate feature engineering approaches, which is a big deal for foundation models <ref:2605.15910#pg1>.
Mira: It gives you an explicit per-grade head for tasks like a CP asymmetry test, where you can explicitly look at the grade four coefficient as that one genuine channel <ref:2605.15910#pg2>.
Lev: And they suggested using an input ablation study to figure out which algebraic content actually earns its place, like quantifying how much structure a reference blade encoding the beam plane carries <ref:2605.15910#pg4>.
Kai: They also mentioned implementing a flavor-flavour universality check by using soft weight-tying as an S3 permutation among e, mu, and tau tokens of the same lepton type <ref:2605.15910#pg4>.
Mira: It’s a common notational tool, but it’s a way to organize what we already measure rather than introducing entirely new physics <ref:2605.15910#pg3>.
Lev: It’s about making sure the model respects Lorentz symmetry without needing to dictate the features itself <ref:2605.15910#pg3>.
Kai: So that's the picture on "Geometric algebra as the input language of collider foundation models." If you want to see how this translates into actual hardware performance, we’ve got another paper about signatures of leptophilic baryogenesis at muon colliders coming up next.
Skobeltsyn Institute of Nuclear Physics, Lomonosov Moscow State University (SINP MSU)
hep-ph
Submitted: 2026-05-15
Updated: 2026-11-18
The gist: A hard hadron-collider event is fundamentally represented as a single geometric object, specifically a graded multivector…
Key concepts
- Graded Multivector
- A single geometric object from geometric algebra ($ ext{Cl}(1, 3)$) representing the entire hard collision event. It organizes physical observables into different 'grades,' where grade zero holds invariants, grade one holds four-momenta, and higher grades represent increasingly complex structures like decay planes.
- Geometric Algebra ($ ext{Cl}(1, 3)$)
- A mathematical framework used to describe the geometry of spacetime. It allows physical quantities like vectors (four-momenta) and bivectors (decay planes) to be combined algebraically. This structure naturally organizes all relevant observables into a consistent hierarchy based on their grade.
- Grade Decomposition Interface
- The fixed organizational structure imposed by the multivector's grade decomposition. This means different downstream tasks can access specific information directly—for instance, a task requiring strict invariance only looks at the grade zero coefficients, while another uses the full object for covariance.
- Cayley–Menger Collapse
- A theorem stating that higher-grade objects built from measured momenta do not introduce new Lorentz invariants beyond what is already encoded in their Gram matrix. This constraint limits how new information can be added to the invariant content of the system.
Terminology
Summary
A hard hadron-collider event is fundamentally represented as a single geometric object, specifically a graded multivector, rather than as a list of four-momenta accompanied by discrete labels for various properties like charge or b-tag information. p. 1 This representation leverages geometric algebra (Cl(1, 3)) whose grade decomposition naturally organizes all relevant observables: grade zero for invariants, grade one for four-momenta, grade two for decay-plane bivectors, and so on up to the CP-odd pseudoscalar at grade four. p. 0
The core proposition is that this single multivector representation serves as a uniform input layer for foundation models in collider physics. p. 0 This framework unifies three previously separate feature engineering approaches. p. 2 The physical meaning of each grade and operation is cataloged, supported by an explicit per-grade dictionary of 32 classical observables and a catalogue of discrete symmetries acting on the event multivector. p. 3
Theorem 1 establishes the constraints on new Lorentz-invariant scalars unlocked by this representation: none are found beyond those already present in the initial set. p. 0 The only genuine non-trivial invariant content that persists is the CP-odd sign of the pseudoscalar (grade four). p. 8 The construction is bound by the Cayley–Menger collapse (Theorem 1), which dictates how higher-grade objects built from measured momenta contribute to Lorentz invariants—they add no new information beyond what their Gram matrix already encodes. p. 8 Crucially, a grade- k object encoding structure that the momenta do not contain—such as the laboratory frame singled out by the beam—does carry structure that is preserved. p. 0
The representation dictates an input specification for a foundation model program (HiGEN), where the event multivector E is used as the per-token input for a Lorentz-equivariant model pre-trained across different processes. p. 3 The architectural details of the encoder and pre-training strategies are left to future work, but the framework provides a structured method for incorporating kinematic and symmetry information directly into neural network inputs. p. 2
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Input representation for foundation models should be a single geometric object, one graded multivector E, rather than a list of four-momenta with discrete labels attached to each object. This allows the network to respect Lorentz symmetry through its algebraic structure without needing auxiliary machinery for every layer.
-
The grade decomposition of the event multivector provides a fixed grade interface for every downstream task, meaning pre-training on E preserves grade structure end-to-end, allowing tasks requiring strict invariance to read grade zero coefficients while those requiring full covariance read the full multivector.
-
The system can be designed to use explicit per-grade heads for fine-tuning classification and regression tasks; for example, a CP asymmetry test could explicitly look at the grade four coefficient, which is the
one genuine channel.
-
The system can utilize an input ablation study to determine which algebraic content earns its place; for instance, it can quantify that
a fixed reference blade encoding the beam plane carries structure the momenta do not contain
by comparing performance against different input representations. -
The system can implement a flavor-flavour universality check as a soft weight-tying mechanism on feature channels, which is realized as an S3 permutation among e, mu, and tau tokens of the same lepton-type label.
Sources
- General recipe to form input space for deep learning analysis of HEP scattering processes
- Lorentz Group Equivariant Neural Network for Particle Physics
- PELICAN: Permutation Equivariant and Lorentz Invariant or Covariant Aggregator Network for Particle Physics
- Geometric Algebra Transformer
- Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics
- A Lorentz-Equivariant Transformer for All of the LHC
- Euclidean, Projective, Conformal: Choosing a Geometric Algebra for Equivariant Transformers
- Realizability of the Lorentzian (n,1)-Simplex
- Solving Key Challenges in Collider Physics with Foundation Models
- Explainable Equivariant Neural Networks for Particle Physics: PELICAN
- An Efficient Lorentz Equivariant Graph Neural Network for Jet Tagging
- Clifford Group Equivariant Neural Networks
- E(n) Equivariant Graph Neural Networks
- Energy Flow Networks: Deep Sets for Particle Jets
- ParticleNet: Jet Tagging via Particle Clouds
- Particle Transformer for Jet Tagging
- Jet Tagging with More-Interaction Particle Transformer
- Permutationless Many-Jet Event Reconstruction with Symmetry Preserving Attention Networks
- SPANet: Generalized Permutationless Set Assignment for Particle Physics using Symmetry Preserving Attention
- Top quark spin correlations and polarization at the LHC: standard model predictions and effects of anomalous top chromo moments
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses