Machine-learned particle flow as a foundation model for collider physics
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Machine-learned particle flow as a foundation model for collider physics".
Mira: Here's the gist of the paper. Machine-learned particle flow as a foundation model for collider physics shows that casting event reconstruction as a machine learning problem naturally produces such a shared representation Foundation Model Concept The paper demonstrates that the latent representations learned during particle-flow reconstruction encode essential physics information needed for downstream analysis, establishing MLPF as a foundation model and offering a concrete step toward an end-to-end pipeline from detector data to physics analysis The authors repurpose a machine learning model trained for particle-flow reconstruction (MLPF) to perform three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression They further demonstrate that a single linear layer trained using only the latent representations achieves competitive performance against state-of-the-art baseline architectures This suggests that MLPF already produces representations that are generically useful for downstream analysis tasks How the training signal is built The MLPF algorithm was originally designed as a graph neural network (GNN) and later evolved into a transformer-based architecture Per-particle latent representations are extracted from the layer immediately preceding the output heads and are then compressed using unsupervised principal component analysis (PCA) to retain only 128 dimensions This compression step is task-agnostic, demonstrating that the MLPF latent representations carry generic information useful across different downstream tasks Where the gains come from The performance improvements are observed across three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression For jet flavor identification, the Latent-augmented model dominates both AUC scaling curves at every training-set size The Linear-probe model results strengthen the foundation-model interpretation by quantifying the fraction of task-relevant information that is already linearly accessible in the MLPF latent space For missing momentum regression, a 129-parameter linear layer outperforms the DeepMET-based Baseline model at every training-set size while using approximately 35 times fewer parameters Pruning redundant reasoning tokens The results confirm that the performance gains observed for the Latent-augmented model in the paper results are not a consequence of the increased parameter count, but reflect genuine additional information supplied by the MLPF latent representations For jet flavor identification, both models continue to improve with dataset size, yet the gap between them persists The authors suspect that the global attention mechanism in MLPF encodes the collective neutral deposits from the full cascade jointly, a signal that is absent from the per-particle PF candidate interface and that the downstream model cannot reconstruct from kinematic features alone Combined with cross-detector transfer-learning studies, this work establishes a two-axis foundation-model picture for machine learning based event reconstruction Reconstruction-level latent representations therefore reduce the task-specific need for feature engineering A similar implication holds for missing momentum regression, where an estimator built on MLPF latent representations could improve performance at fixed data scale with no modification to the upstream reconstruction Acknowledgments The authors acknowledge the use of Anthropic’s Claude (Opus 4.7) for assistance with code development and manuscript editing F.M.
Kai: and J.D.
Mira: are supported by the Research Corporation for Science Advancement (RCSA) under grant No DE-SC0009919, and the U.S.
Kai: National Science Foundation (NSF) Harnessing the Data Revolution (HDR) Institute for Accelerating AI Algorithms for Data Driven Discovery (A3D3) under Cooperative Agreement No PHY-2117997 F.M.
Title and authors: Kai: So we're looking at this paper titled "Machine-learned particle flow as a foundation model for collider physics". It seems to be about taking something that's traditionally a separate reconstruction step and using it as a base for other things.
Mira: Right, it looks like they show how casting event reconstruction as a machine learning problem naturally produces this shared representation, which is the central idea here. This suggests that the way we reconstruct events itself generates information useful for different physics analyses without us having to explicitly train those analysis models from scratch on new data.
Lev: From an engineering standpoint, it's interesting because they repurpose a machine learning model trained just for particle flow reconstruction, the MLPF model, to do three separate tasks: jet flavor identification, jet energy regression, and missing momentum regression. That’s a big step in terms of how we reuse existing models.
Kai: Exactly. And what they show is that by just appending those per-particle latent representations learned during reconstruction as extra input features to these downstream tasks, you get substantially better performance compared to baselines that only use standard kinematic features.
Mira: That's the core claim: these MLPF latent representations are generically useful for different physics analysis tasks, meaning the model backbone doesn't need modification or further training for those specific analyses.
Lev: And they do show a very efficient way to leverage this information, they demonstrate that a single linear layer trained using only those latent representations can actually achieve competitive performance against state-of-the-art architectures.
Kai: That’s the part where I think it gets really interesting because it suggests that the MLPF latent space already encodes essential physics information needed for downstream analysis, which is pretty powerful if true.
Mira: It seems like they quantify this by showing how much task relevant information is linearly accessible in that latent space, which helps us understand if the backbone already contains specific structures we need to find.
Lev: And for missing momentum regression specifically, they show that a linear layer with only one hundred twenty-nine parameters can beat the DeepMET-based Baseline model at every training set size while using about thirty-five times fewer parameters. That efficiency is notable when you think about running things on real hardware.
Kai: So it’s not just that it works better, it’s that the way they've done this suggests we might not need to treat ML-based reconstruction and physics analysis as completely separate pipeline stages anymore.
Mira: That’s what the paper points toward: reconstruction models can serve as natural foundation models for collider physics, which means we don't have to engineer those features from scratch every time.
Lev: It sounds like they found a way to reduce that task-specific need for feature engineering by just using these reconstruction-level latent representations.
Kai: So, before we wrap up this discussion on "Machine-learned particle flow as a foundation model for collider physics", we've talked about how the MLPF model is used across jet flavor identification, energy regression, and missing momentum regression.
Mira: And what they're really highlighting is that the latent representations are what connect those tasks together without needing any modification to the original reconstruction pipeline.
Lev: I think it’s important to remember that one limitation they point out is that their results are based on simulated datasets from the Compact Linear Collider detector, and while they finetuned on CLD data, you still need those simulations for testing.
Kai: Right, so it’s a strong result in terms of showing the potential structure in the representations, even if the current validation relies heavily on that simulation setup.
Mira: It really does suggest a path forward where we can leverage these reconstruction models as natural foundation models for our future collider physics work.
Lev: I just think it’s a great demonstration of how to get meaningful physics insights out of the raw reconstruction process itself.
Kai: Exactly. So, that's what we have on "Machine-learned particle flow as a foundation model for collider physics". It shows us how to build these shared representations and use them for things like jet flavor identification, energy regression, and missing momentum regression without needing extra training on the backbone itself.
The paper's summary: Kai: So, to recap, this paper argues that you can take a machine learning model built just for reconstructing particle flow in events and reuse its internal structure—the latent representations—for totally different physics tasks like identifying jet flavor or figuring out jet energy.
Mira: Right, it’s about establishing that particle flow reconstruction isn't just a black box for finding the final state; it actually produces a really useful internal language that other models can speak to directly.
Lev: From my side, what I find compelling is how they show you can train a single linear layer using only those latent representations and still get better results than the complex, high-parameter architectures you usually build for these analyses.
Kai: It’s not just that it works; it seems like the paper quantifies exactly how much of that useful information is already there, meaning you don't have to spend as much effort engineering those features yourself.
Mira: That’s what I mean, because they show a linear probe model that tells you how much task relevant information is actually accessible in that latent space, independent of any complicated layers the downstream model might use.
Lev: For missing momentum regression specifically, they showed a linear layer with just one hundred twenty-nine parameters beating a deep meta-model baseline by thirty-five times less parameter count. That level of efficiency on real hardware is what’s important for me.
Kai: So, the main implication here is that we can stop treating event reconstruction and physics analysis as two totally separate training problems that need to be perfectly matched up later.
Mira: It suggests a unified foundation model picture where the reconstruction step provides the universal representation for everything else in the detector data, which really reduces task-specific feature engineering.
Lev: If we can use these representations directly, it means we don't need to modify our existing reconstruction pipeline just because we want to do a new physics analysis.
Kai: So it’s about building this two-axis foundation model picture for event reconstruction, where the representation level is the key connector between everything.
Mira: And this connection works because the latent representations learned during particle flow naturally capture those essential physics nuances needed for things like jet flavor identification.
Lev: We've seen how promising it is, but we still need to be careful about what that reconstruction model actually learned; it’s based on specific detector data, which we have to keep in mind.
Kai: Exactly, so while the results are strong on the simulation side, the real test will be how robust these representations are when we finally run them on actual detector data.
Mira: That’s what we need to look at next: how does this foundational representation transfer when you move from one kind of event to another?
The paper's improvements: Kai: So, we’re looking at how they suggest improving the model itself beyond just using the latent representations for analysis, right?
Mira: Yeah, they show that you can actually use a linear layer trained on those representations to get competitive performance against these more complex architectures. It’s about showing that simplicity can be powerful here.
Lev: What I see is them quantifying that linear probe model so you can actually measure how much task relevant information is accessible in the latent space, even when you aren't doing any nonlinear processing downstream.
Kai: So it’s a diagnostic tool, helping us understand if the reconstruction backbone already has those specific physics structures encoded within its representations.
Mira: That helps pin down the assumptions; if that linear probe shows high information density in a specific area, then the assumption that particle flow reconstruction produces useful general features holds up for that task.
Lev: For missing momentum regression, they used this idea to show a tiny one hundred twenty-nine parameter linear layer beating a deep meta model by thirty-five times less parameter count while maintaining performance across all training set sizes.
Kai: That’s a really strong number showing how much complexity you can strip away and still get good results for that specific task.
Mira: It means that for certain regressions, we don't need to throw in massive neural networks if the underlying latent structure is already doing the heavy lifting linearly.
Lev: But I gotta say, they flag that this works best on simulated datasets from something like the Compact Linear Collider detector, so it’s not quite a guaranteed plug-and-play for real hardware yet.
Kai: Right, so we have these strong results in terms of showing the potential structure, but we still need to see how robust those representations are when we run them on actual detector data.
Mira: That brings up the next big question: if this foundation model approach works for jet flavor identification, does that same mechanism apply when you look at other types of physics signals in these collisions?
Conclusion: Tom: So we’re wrapping up this look at "Machine-learned particle flow as a foundation model for collider physics," and it boils down to this: they show that you can take the particle flow reconstruction process itself and use its internal structure as a base for many other physics analyses.
Kai: Basically, the paper proves that casting event reconstruction as a machine learning problem naturally generates those shared representations we need for downstream tasks like jet flavor identification or missing momentum regression.
Mira: Right, it’s about establishing that particle flow reconstruction isn't just a black box for finding the final state; it actually produces a really useful internal language that other models can speak to directly.
Lev: From my side, what I find compelling is how they show you can train a single linear layer using only those latent representations and still get competitive performance against these more complex architectures.
Kai: It’s not just that it works; it seems like the paper quantifies exactly how much of that useful information is already there, meaning you don't have to spend as much effort engineering those features yourself.
Mira: That helps pin down the assumptions; if that linear probe shows high information density in a specific area, then the assumption that particle flow reconstruction produces useful general features holds up for that task.
Lev: For missing momentum regression specifically, they used this idea to show a tiny one hundred twenty-nine parameter linear layer beating a deep meta model baseline by thirty-five times less parameter count while maintaining performance across all training set sizes.
Kai: That’s a really strong number showing how much complexity you can strip away and still get good results for that specific task.
Mira: It means that for certain regressions, we don't need to throw in massive neural networks if the underlying latent structure is already doing the heavy lifting linearly.
Lev: But I gotta say, they flag that this works best on simulated datasets from something like the Compact Linear Collider detector, so it’s not quite a guaranteed plug-and-play for real hardware yet.
Kai: Right, so we have these strong results in terms of showing the potential structure, but we still need to see how robust those representations are when we run them on actual detector data.
Mira: That brings up the next big question: if this foundation model approach works for jet flavor identification, does that same mechanism apply when you look at other types of physics signals in these collisions?
Lev: I just think it’s a great demonstration of how to get meaningful physics insights out of the raw reconstruction process itself.
Kai: So, the main implication here is that we can stop treating event reconstruction and physics analysis as two totally separate training problems that need to be perfectly matched up later.
Mira: It suggests a unified foundation model picture where the reconstruction step provides the universal representation for everything else in the detector data, which really reduces task-specific feature engineering.
Lev: We've seen how promising it is, but we still need to be careful about what that reconstruction model actually learned; it’s based on specific detector data, which we have to keep in mind.
Kai: Exactly, so while the results are strong on the simulation side, the real test will be how robust these representations are when we run them on actual detector data.
Mira: That really shows how powerful this paper is for guiding future model development in high energy physics analysis.
Lev: I just think it’s a great demonstration of how to get meaningful physics insights out of the raw reconstruction process itself.
University of California San Diego · National Institute of Chemical Physics and Biophysics · SLAC National Accelerator Laboratory
hep-ex, cs.LG, hep-ph, physics.data-an, physics.ins-det
Submitted: 2026-06-12
Updated: 2026-11-04
Comments: 16 pages, 14 figures
Journal ref: PRX INTELLIGENCE 1, 023012 (2026)
DOI: 10.1103/v5n3-d3tn
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Machine-learned particle flow as a foundation model for collider physics shows that casting event reconstruction as a machine learning problem naturally produces such a shared representation…
Key concepts
- Foundation Model Concept
- The MLPF algorithm is treated as a foundation model because the latent representations learned during particle flow reconstruction contain core physical information. These representations are generic enough to be used directly for multiple, different physics analysis tasks without needing task-specific retraining or modification of the original model.
- Latent Representations
- These are compressed, low-dimensional numerical summaries extracted from the MLPF layers. They capture essential physics details about a particle's path and interaction within the detector. Crucially, these representations are task-agnostic, meaning they hold general physical knowledge applicable across different analysis goals.
- Linear Probe Model
- This technique quantifies how much of the information relevant to a specific downstream task is already linearly accessible within the MLPF latent space. It acts as a diagnostic tool to measure the inherent structure of the learned representations, showing which physics signals are immediately usable by simple linear models.
- Task-Agnostic Compression
- The process of compressing particle-flow latent representations down to 128 dimensions using unsupervised PCA is task-agnostic. This means the resulting compressed features retain generic physical information that is useful across different analysis goals, rather than being tailored for just one specific output.
Terminology
Summary
Machine-learned particle flow as a foundation model for collider physics shows that casting event reconstruction as a machine learning problem naturally produces such a shared representation
Foundation Model Concept
The paper demonstrates that the latent representations learned during particle-flow reconstruction encode essential physics information needed for downstream analysis, establishing MLPF as a foundation model and offering a concrete step toward an end-to-end pipeline from detector data to physics analysis The authors repurpose a machine learning model trained for particle-flow reconstruction (MLPF) to perform three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression They further demonstrate that a single linear layer trained using only the latent representations achieves competitive performance against state-of-the-art baseline architectures This suggests that MLPF already produces representations that are generically useful for downstream analysis tasks
How the training signal is built
The MLPF algorithm was originally designed as a graph neural network (GNN) and later evolved into a transformer-based architecture Per-particle latent representations are extracted from the layer immediately preceding the output heads and are then compressed using unsupervised principal component analysis (PCA) to retain only 128 dimensions This compression step is task-agnostic, demonstrating that the MLPF latent representations carry generic information useful across different downstream tasks
Where the gains come from
The performance improvements are observed across three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression For jet flavor identification, the Latent-augmented model dominates both AUC scaling curves at every training-set size The Linear-probe model results strengthen the foundation-model interpretation by quantifying the fraction of task-relevant information that is already linearly accessible in the MLPF latent space For missing momentum regression, a 129-parameter linear layer outperforms the DeepMET-based Baseline model at every training-set size while using approximately 35 times fewer parameters
Pruning redundant reasoning tokens
The results confirm that the performance gains observed for the Latent-augmented model in the paper results are not a consequence of the increased parameter count, but reflect genuine additional information supplied by the MLPF latent representations For jet flavor identification, both models continue to improve with dataset size, yet the gap between them persists The authors suspect that the global attention mechanism in MLPF encodes the collective neutral deposits from the full cascade jointly, a signal that is absent from the per-particle PF candidate interface and that the downstream model cannot reconstruct from kinematic features alone
Conclusion
Combined with cross-detector transfer-learning studies, this work establishes a two-axis foundation-model picture for machine learning based event reconstruction Reconstruction-level latent representations therefore reduce the task-specific need for feature engineering A similar implication holds for missing momentum regression, where an estimator built on MLPF latent representations could improve performance at fixed data scale with no modification to the upstream reconstruction
Acknowledgments
The authors acknowledge the use of Anthropic’s Claude (Opus 4.7) for assistance with code development and manuscript editing F. p. 4 For
charged PF candidates, we therefore additionally derive
the full set of displacement variables listed in Table III,
following Ref. [35] and the numerical implementation of
the official CLD flavor-tagging code [36]: the transverse
and longitudinal IPs (d0, z0), the signed 2D and 3D
IPs and their significances (SIP2D, SIP2D/σ2D, SIP3D,
SIP3D/σ3D), the 3D distance between the track and the
jet axis at their point of closest approach (d3D) and its
significance (d3D/σd3D ), and the full 15-element track
parameter covariance matrix Cij . Model variants for the jet flavor identification
task. The Linear-probe model is a single linear layer with
387 parameters (128 × 3 weights + 3 biases), where the factor
of 3 corresponds to the three flavor classes (b, c, light)
Models are trained on a heterogeneous pool of GPUs
(NVIDIA A10, L40S, V100, RTX 3090, and RTX 4090)
on the National Research Platform (NRP) Kubernetes
cluster [39], with wall-clock training time ranging from
∼14 minutes (smallest dataset, fastest GPU) to ∼9 hours
(full dataset, on slower GPUs). Model variants for the jet flavor identification
task. The Linear-probe model is a single linear layer with
387 parameters (128 × 3 weights + 3 biases), where the factor
of 3 corresponds to the three flavor classes (b, c, light)
Models are trained on a heterogeneous pool of GPUs
(NVIDIA A10, L40S, V100, RTX 3090, and RTX 4090)
on the National Research Platform (NRP) Kubernetes
cluster [39], with wall-clock training time ranging from
∼14 minutes (smallest dataset, fastest GPU) to ∼9 hours
(full dataset, on slower GPUs). Machine-learned particle flow as a foundation model for collider physics
Farouk Mokhtar ,
1, ∗ Joosep Pata ,
2 Michael Kagan ,
3 and Javier Duarte 1
1University of California San Diego, La Jolla, California 92093, USA
2National Institute of Chemical Physics and Biophysics, Tallinn 12618, Estonia
3SLAC National Accelerator Laboratory, Menlo Park, California 94025, USA
The workflow from particle collision to physics analysis passes through a series of reconstruction
steps that are traditionally modular and disconnected, with no shared representation linking low-level
detector data to high-level analysis tasks. We show that casting event reconstruction as a machine
learning problem naturally produces such a shared representation.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
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
-
Foundational Model for Event Reconstruction: The MLPF model can be repurposed to perform three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression. This establishes MLPF as a foundation model that produces shared representations useful for downstream physics analysis without requiring explicit task-specific training or modification of the backbone.
-
Cross-Task Transfer Capability: The latent representations learned during reconstruction can be appended to standard PF candidate kinematic features to substantially improve performance across different analysis tasks, such as jet flavor identification, which requires track-level quantities like impact parameters. This demonstrates that
reconstruction models may serve as natural foundation models for collider physics.
-
Parameter Efficient Analysis: A single linear layer trained using only the latent representations achieves competitive performance against state-of-the-art architectures while outperforming the baseline for missing momentum regression with approximately 35 times fewer parameters. This suggests that
latent representations learned during reconstruction encode essential physics information needed for downstream analysis.
-
Feature Engineering Reduction: For jet flavor identification, a 387-parameter linear classifier using only MLPF latent representations achieves non-trivial discrimination without any hand-crafted track IP features. This reduces the task-specific need for feature engineering by leveraging
reconstruction-level latent representations.
-
Dynamic Information Quantification: The Linear-probe model can quantify
how much task relevant information is linearly accessible in the latent space, independently of any nonlinear processing by the downstream architecture.
This diagnostic tool helps researchers understand if the MLPF backbone already encodes specific physics structures.
Abstract
The workflow from particle collision to physics analysis passes through a series of reconstruction steps that are traditionally modular and disconnected, with no shared representation linking low-level detector data to high-level analysis tasks. We show that casting event reconstruction as a machine learning problem naturally produces such a shared representation. We repurpose a machine learning model trained for particle-flow reconstruction (MLPF) to perform three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression. By appending the per-particle latent representations learned during reconstruction as additional input features, we substantially improve over baselines that use conventional particle-level features alone. We further demonstrate that a single linear layer trained using only the latent representations achieves competitive performance against state-of-the-art baseline architectures, and outperforms the baseline for missing momentum regression with approximately 35 times fewer parameters. These results demonstrate that the latent representations learned during reconstruction encode essential physics information needed for downstream analysis, establishing MLPF as a foundation model and offering a concrete step toward an end-to-end pipeline from detector data to physics analysis.
Sources
- Particle-flow reconstruction and global event description with the CMS detector
- MLPF: Efficient machine-learned particle-flow reconstruction using graph neural networks
- On the Opportunities and Risks of Foundation Models
- Masked Particle Modeling on Sets: Towards Self-Supervised High Energy Physics Foundation Models
- OmniJet-$\alpha$: The first cross-task foundation model for particle physics
- Is Tokenization Needed for Masked Particle Modelling?
- Re-Simulation-based Self-Supervised Learning for Pre-Training Foundation Models
- Pre-training strategy using real particle collision data for event classification in collider physics
- Pretrained Event Classification Model for High Energy Physics Analysis
- Foundation model for particle collision data analysis
- Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics
- Finetuning Foundation Models for Joint Analysis Optimization
- Accelerating Resonance Searches via Signature-Oriented Pre-training
- Large-Scale Pretraining and Finetuning for Efficient Jet Classification in Particle Physics
- Foundation models for high-energy physics
- Machine Learning
- Improved particle-flow event reconstruction with scalable neural networks for current and future particle detectors
- Fine-tuning machine-learned particle-flow reconstruction for new detector geometries in future colliders
- End-to-end event reconstruction for precision physics at future colliders
- Full event interpretation with machine-learning-based particle-flow reconstruction in the CMS detector
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques