AlphaHits: Full-Event Tracking and Reconstruction Towards a Hit-Level Foundation Model
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: We're looking at the paper: "AlphaHits: Full-Event Tracking and Reconstruction Towards a Hit-Level Foundation Model".
Mira: Here's the gist of the paper. AlphaHits introduces AlphaHits, a Transformer-based approach towards hit-level foundation models for LHC physics building on the TrackFormer and HiggsFormer architectures The model demonstrates scalable full-event charged-particle reconstruction together with multi-task event reconstruction and classification from low-level detector information.
Kai: Pruning redundant reasoning tokens AlphaHits operates directly on detector hits and demonstrates scalable full-event charged-particle reconstruction together with multi-task event reconstruction and classification from low-level detector information AlphaHits is pretrained on charged-particle tracking and fine-tuned for multi-task reconstruction and classification from a shared hit-level representation The model processes both pixel and strip detector hits in events containing up to 1.5 × 10 5 hits under HL-LHC-like pileup conditions AlphaHitsTrackFormer reconstructs the full charged-particle content of complete TrackML events without imposing a transverse-momentum selection on the target tracks AlphaHits achieves a TrackML score of 0.959 when evaluated inclusively over all reconstructable tracks representing, to our knowledge, the highest reported TrackML score for full-detector, inclusive charged-particle reconstruction Building on the TrackFormer architecture, AlphaHits is pretrained on charged-particle tracking and fine-tuned for multi-task reconstruction and classification from a shared hit-level representation From a shared hit-level representation, it predicts tracks, jets, jet flavour, and event-level observables How the training signal is built AlphaHits utilizes two complementary stages for training: Stage 1 trains the backbone as a Tracker from scratch using only tracking and hit-particle heads Stage 2 adds the other seven heads with fresh weights and trains them jointly, warm-started from the Stage 1 checkpoint The Stage 2 loss is a geometric mean over the six main tasks plus fixed-weight auxiliary terms The main tasks include event class, HT, E miss T, leptons, jets set (12 jet queries + flavour), and tracking token prediction Auxiliary terms are added with fixed weights outside the product to ensure a near-zero scaffold loss cannot collapse it Where the gains come from The performance gains stem from the combination of scalable full-event tracking with multi-task learning from shared hit-level representations AlphaHits achieves an AUC of 0.944 for Higgs event classification, demonstrating performance comparable to that of a Particle Transformer (ParT) baseline using fully reconstructed inputs The results point towards a common paradigm in which scalable sparse neural networks learn physically informative representations directly from detector measurements AlphaHits extends this approach to multi-task event reconstruction and classification exploring whether a common representation of low-level detector information can be reused across reconstruction and inference tasks Performance comparison against baselines AlphaHits achieves a TrackML score of 0.959 over all tracks (pT > 0.15 GeV) on the TrackML dataset AlphaHits reaches an AUC of 0.932 for Higgs event classification from silicon hits alone and 0.944 with calorimeter towers as an additional input The results show that AlphaHits finds all six GenJets with the right flavour when using hits only, matching the Delphes track-jet reference from 77 % b-efficiency on (light-jet rejection ratio 1.0, 1.0 and 1.3 at 77, 80 and 85 %; c-jet 1.0, 1.1 and 1.3) Together, these results establish scalable full-event tracking and multi-task learning as complementary steps towards general-purpose hit-level foundation models and Large Physics Models With larger-scale training and further developments, such approaches could ultimately enable end-to-end optimization of physics analyses directly from detector measurements potentially improving sensitivity beyond conventional reconstruction pipelines The objective is to optimize the analysis chain end-to-end for specific physics goals potentially surpassing the performance of conventional reconstruction pipelines and enabling sensitivity to signals that are currently difficult or impossible to reconstruct The focus of this work was the development of hit-level tracking and foundation models and their physics performance The one-shot model keeps the tracking head but reaches a TrackML score of 0.806 with the equal-token decode and a double-majority efficiency of 0.893 for pT > 1 GeV The towers improve the tracking (0.819, 0.897) The AlphaHits backbone is smaller than the AlphaHits-TrackFormer Tracker (12 instead of 18 layers, 9.7 M instead of 19.4 M parameters) and is trained on a different sample with a different codebook so the two tracking scores are not directly comparable The depth dependence of Figure 9 suggests that a deeper backbone would also lead to better tracking performance The authors acknowledge the assistance of large language models, including GPT-5, GPT-6, and Claude Opus for code generation, debugging, and implementation support The study builds on previous implementations of TrackFormer and HiggsFormer which were developed with the assistance of large language models including GPT-5, GPT-6, and Claude Opus for code generation The work was supported by the European Union ERDF and Comunitat Valenciana as well as technical support provided by the Instituto de F´ısica Corpuscular, IFIC (CSIC-UV) The paper is 2611.13222 AlphaHits-TrackFormer therefore processes the full pixel and strip hit cloud and attempts to reconstruct all reconstructable charged particles represented in the TrackML sample without imposing a pT threshold on the target tracks We further ask whether such hit-level representations can support tasks beyond charged-particle tracking AlphaHits extends this approach to multi-task event reconstruction and classification exploring whether a common representation of low-level detector information can be reused across reconstruction and inference tasks Page 6 --- Our earlier TrackFormer and HiggsFormer studies explored this idea from complementary directions TrackFormer investigated Transformer-based assignment of inner-tracker hits to charged-particle trajectories HiggsFormer showed that a related architecture could be retrained to classify Higgs events directly from raw inner-tracker hits bypassing explicit track and object reconstruction Trained on the raw inner-tracker hits of simulated ttH¯ (H → b¯b) and tt¯ events, it reached AUC of 0.855, comparable to what a Particle Transformer
5: on Delphes tracker objects reaches at the 40% b-tagging working point (0.852) This result suggested that models trained on low-level inputs can learn detector-hit representations which retain substantial information relevant not only to particle reconstruction but also to downstream physics inference Page 7 --- AlphaHits-TrackFormer reconstructs tracks in complete TrackML events of up to 1.5 × 10 5 hits A learned Router sparsifies the attention a global Tracker model predicts binned and continuous track parameters for every hit a charge-aware preconditioner and a learned Gate select the ambiguous regions and a local Refiner resolves them AlphaHits-TrackFormer reaches TrackML scores of 0.959 over all tracks (pT > 0.15 GeV) trained on 7830 events Page 8 --- The AlphaHits-TrackFormer reconstruction pipeline is hierarchical A learned Router first assigns detector hits to overlapping logical groups that determine the sparse attention structure A full-event Tracker then predicts track parameters for every hit and maps the hits into a discrete track-parameter space Finally, a learned Gate identifies regions that require additional processing and a compact local Refiner resolves the remaining combinatorial ambiguities Page 9 --- The Router is a small MLP that processes hits independently It receives the continuous hit features together with learned embeddings describing the detector volume layer and module and predicts membership in two of four × 4 × 32 groups made in hit-radius track-pseudorapidity and track-azimuth The Router therefore serves two purposes it makes full-event attention computationally tractable and introduces a learned locality structure determining which detector regions can communicate directly Page 10 --- Hits, used as an input are represented as short vectors of their position and local geometry They are embedded by one linear layer into d=256 vectors and passed through a stack of Transformer encoder layers The output of the encoder is one state per hit and every head of both models reads these per-hit states AlphaHits adds learned embeddings of the detector volume layer and module and eight cluster-shape features for every input hit Page 11 --- AlphaHits predicts, for every hit its particle’s kinematics, production-vertex displacement and charge, and its member jet’s axis pT and secondary-vertex flight distance next to the track token The auxiliary terms enter the loss with fixed weights outside the main objective Page 12 --- AlphaHits predicts for every hit its particle’s kinematics, production-vertex displacement and charge, and its member jet’s axis pT and secondary-vertex flight distance next to the track token The auxiliary terms enter the loss with fixed weights outside the main objective Page 13 --- AlphaHits is a permutation-equivariant Transformer architecture operating directly on detector hits building on the TrackFormer and HiggsFormer models The first HiggsFormer was an event classifier over raw hit coordinates (x, y, z) with a single ttH¯ (H → b¯b) vs tt¯ head plus an auxiliary HT regression head for the ‘big’ model Page 14 --- Each hit is the seven-dimensional vector: position in metres transverse radius azimuth as (cos φ,sin φ) and pseudorapidity The model sees positions only no detector identifiers cluster shape charge or timing Page 15 --- AlphaHits is a permutation-equivariant Transformer architecture operating directly on detector hits building on the TrackFormer and HiggsFormer models It is pretrained on tracking tasks only (Stage 1) and fine-tuned with nine heads to predict tracks, jets with flavour, HT, Emiss T, leptons and the event class in one forward pass (Stage 2) Page 16 --- The first HiggsFormer was an event classifier over raw hit coordinates (x, y, z) with a single ttH¯ (H → b¯b) vs tt¯ head plus an auxiliary HT regression head for the ‘big’ model AlphaHits builds on the tracking backbone of Section 4 Page 17 --- All training of the AlphaHits-TrackFormer models was conducted on an NVIDIA H200 card and took roughly 1 week for the best model The Router is first pretrained for one epoch using only the supervised routing objective The Router and Tracker are subsequently trained jointly for 50 epochs using the AdamW optimizer with an initial learning rate of 3 × 10-4, weight decay of 1 × 10-4, and cosine learning-rate annealing Page 22 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 24 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 26 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 27 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 28 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 29 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 30 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 31 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 32 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 33 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 34 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 35 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 36 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 37 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 38 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 39 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 40 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 41 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 42 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 43 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 44 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 45 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 46 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 47 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 48 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 49 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision Page 50 --- The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
Paper summary: Kai: So, looking at AlphaHits: Full-Event Tracking and Reconstruction Towards a Hit-Level Foundation Model, the main thrust is that using this Transformer approach directly on detector hits allows for scalable full event charged particle reconstruction alongside multi-task learning from those low level inputs.
Mira: It’s about proving that you can reuse a common representation of low level detector information across different physics objectives, which they explore through event reconstruction and classification tasks.
Lev: In simple terms, this means we might be able to optimize the entire analysis chain end-to-end by starting directly from the detector measurements instead of relying on a fixed pipeline.
Kai: The title emphasizes that it's about moving towards a hit level foundation model, suggesting these models could eventually allow for end-to-end optimization of physics analyses directly from detector measurements.
Mira: It suggests that with larger scale training and further developments, these approaches could potentially enable sensitivity beyond conventional reconstruction pipelines by optimizing the analysis chain specifically for particular physics goals.
Lev: The paper focuses on developing hit level tracking and foundation models and their performance in physics applications, which is a significant step toward that goal.
Conclusion: Kai: So we've seen how AlphaHits uses this Transformer to reconstruct full events directly from detector hits, and now we're at the conclusion of this paper about AlphaHits: Full-Event Tracking and Reconstruction Towards a Hit-Level Foundation Model.
Mira: Basically, they're saying that by training this model to do tracking and other tasks all at once from those raw hits, you can create a foundation model that works directly on detector measurements.
Lev: It seems like the authors are pushing the idea that this hit-level representation is useful not just for one thing, but for many different physics goals, like reconstruction and event classification.
Kai: Right, they're showing how they build this from scratch using two stages of training—first learning tracking on its own, then adding all those other heads to learn everything jointly.
Mira: And the results are pretty solid on the performance numbers they report for things like Higgs event classification, even when using only silicon hits initially.
Lev: But I always want to ask about how much overhead this model would actually take to run on real hardware, especially when you're dealing with pileup conditions as high as those at the HL-LHC.
Kai: Well, the paper mentions they tested it under conditions up to one point five times ten thousand hits and that's what they focused on for their main reconstruction test.
Mira: The implication here is that if you can learn these powerful representations from just the detector data, you might be able to optimize your entire analysis chain end-to-end.
Lev: That's a big change because it could mean moving away from those fixed reconstruction pipelines and tuning everything directly for a specific physics goal.
Kai: Exactly, they suggest that with bigger training sets and more work on this kind of approach, we might be able to improve sensitivity in ways that are currently just out of reach.
Mira: So it really boils down to whether a shared representation of low level detector information can truly be reused across both reconstruction and inference tasks.
Lev: That's the core question they're leaving open for future work, which is where the real engineering challenge will come in figuring out how to scale this up.
Sascha Caron, Nadezhda Dobreva, Polina Moskvitina, †Eugene Shalugin, Javad Sohrabi, Roberto Ruiz de Austri, ũue Zhao
National Institute for Subatomic Physics, The Netherlands · High-Energy Physics, Radboud University, The Netherlands · CERN, Switzerland · Instituto de Física Corpuscular, CSIC-Universitat de Valáncia · SURF
hep-ex, hep-ph
Submitted: 2026-11-09
Updated: 2026-11-09
Comments: 29 pages, 13 figures, 13 tables
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 88/100
The gist: AlphaHits introduces AlphaHits, a Transformer-based approach towards hit-level foundation models for LHC physics building on the TrackFormer and HiggsFormer architectures The model demonstrates…
Key concepts
- Hit-Level Foundation Models
- These are large AI models trained directly on raw detector hits instead of reconstructed particles. AlphaHits uses this approach to learn general physical representations from low-level data, which can then be reused for various physics tasks.
- TrackFormer Architecture
- This is the underlying Transformer structure used by AlphaHits. It is adapted to process detector hits, allowing it to perform assignments of inner-tracker hits to charged-particle trajectories, similar to how traditional trackers work but using a neural network.
- Multi-Task Learning from Shared Representation
- AlphaHits trains one backbone that learns a common representation from raw hits. This single representation is then used by multiple specialized heads to perform different tasks simultaneously, such as reconstructing tracks and classifying events.
Terminology
Summary
AlphaHits introduces AlphaHits, a Transformer-based approach towards hit-level foundation models for LHC physics building on the TrackFormer and HiggsFormer architectures The model demonstrates scalable full-event charged-particle reconstruction together with multi-task event reconstruction and classification from low-level detector information. p. 0 AlphaHits: Full-Event Tracking and Reconstruction
Towards a Hit-Level Foundation Model
Sascha Caron1,2, Nadezhda Dobreva1,2,3, Polina Moskvitina1,2,†
, Eugene Shalugin1,2,3, Javad
Sohrabi4
, Roberto Ruiz4
, and Yue Zhao5
1National Institute for Subatomic Physics (Nikhef), Amsterdam, The Netherlands
2High-Energy Physics, Radboud University, Nijmegen, The Netherlands
3CERN, Geneva, Switzerland
4
Instituto de F´ısica Corpuscular, CSIC-Universitat de Val`encia, E-46980 Paterna, Valencia,
Spain
5SURF, Amsterdam, The Netherlands
†Current affiliation: Argonne National Laboratory, USA
Abstract
We introduce AlphaHits, a Transformer-based approach towards hit-level foundation models for LHC
physics, building on the TrackFormer and HiggsFormer architectures. AlphaHits operates directly on
detector hits and demonstrates scalable full-event charged-particle reconstruction together with multi-task
event reconstruction and classification from low-level detector information.Where the paper says it
Pruning redundant reasoning tokens
AlphaHits operates directly on detector hits and demonstrates scalable full-event charged-particle reconstruction together with multi-task event reconstruction and classification from low-level detector information AlphaHits is pretrained on charged-particle tracking and fine-tuned for multi-task reconstruction and classification from a shared hit-level representation The model processes both pixel and strip detector hits in events containing up to 1.5 × 10 5 hits under HL-LHC-like pileup conditions AlphaHitsTrackFormer reconstructs the full charged-particle content of complete TrackML events without imposing a transverse-momentum selection on the target tracks AlphaHits achieves a TrackML score of 0. p. 0 AlphaHits operates directly on
detector hits and demonstrates scalable full-event charged-particle reconstruction together with multi-task
event reconstruction and classification from low-level detector information. For charged-particle tracking, AlphaHits-TrackFormer reconstructs the full charged-particle content
of complete TrackML events without imposing a transverse-momentum selection on the target tracks.Where the paper says it
How the training signal is built
AlphaHits utilizes two complementary stages for training: Stage 1 trains the backbone as a Tracker from scratch using only tracking and hit-particle heads Stage 2 adds the other seven heads with fresh weights and trains them jointly, warm-started from the Stage 1 checkpoint The Stage 2 loss is a geometric mean over the six main tasks plus fixed-weight auxiliary terms The main tasks include event class, HT, E miss T, leptons, jets set (12 jet queries + flavour), and tracking token prediction Auxiliary terms are added with fixed weights outside the product to ensure a near-zero scaffold loss cannot collapse it
Where the gains come from
The performance gains stem from the combination of scalable full-event tracking with multi-task learning from shared hit-level representations AlphaHits achieves an AUC of 0.944 for Higgs event classification, demonstrating performance comparable to that of a Particle Transformer (ParT) baseline using fully reconstructed inputs The results point towards a common paradigm in which scalable sparse neural networks learn physically informative representations directly from detector measurements AlphaHits extends this approach to multi-task event reconstruction and classification exploring whether a common representation of low-level detector information can be reused across reconstruction and inference tasks
Performance comparison against baselines
AlphaHits achieves a TrackML score of 0.959 over all tracks (pT > 0. p. 0 AlphaHitsTrackFormer achieves a TrackML score of 0. 959 when evaluated inclusively over all reconstructable
tracks, representing, to our knowledge, the highest reported TrackML score for full-detector, inclusive
charged-particle reconstruction. AlphaHitsTrackFormer achieves a TrackML score of 0. 959 when evaluated inclusively over all reconstructable
tracks, representing, to our knowledge, the highest reported TrackML score for full-detector, inclusive
charged-particle reconstruction. Nine heads are then trained jointly on the same
hits and reconstruct tracks, jets with flavour tags, HT , missing transverse momentum, leptons and the
event class in one forward pass. On the ttH¯ (H → b
¯b) vs tt¯ hard scatter task AlphaHits reaches an AUC
of 0. On the ttH¯ (H → b
¯b) vs tt¯ hard scatter task AlphaHits reaches an AUC
of 0. 932 from silicon hits alone and 0. 944 with calorimeter towers as an additional input. On the same
events, a Particle Transformer on Delphes objects reaches 0. Building on the TrackFormer architecture, AlphaHits is pretrained on charged-particle tracking and
fine-tuned for multi-task reconstruction and classification. From a shared hit-level representation, it
predicts tracks, jets, jet flavour, and event-level observables.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
Together, these results establish scalable full-event tracking and multi-task learning as complementary steps towards general-purpose hit-level foundation models and Large Physics Models With larger-scale training and further developments, such approaches could ultimately enable end-to-end optimization of physics analyses directly from detector measurements potentially improving sensitivity beyond conventional reconstruction pipelines The objective is to optimize the analysis chain end-to-end for specific physics goals potentially surpassing the performance of conventional reconstruction pipelines and enabling sensitivity to signals that are currently difficult or impossible to reconstruct The focus of this work was the development of hit-level tracking and foundation models and their physics performance The one-shot model keeps the tracking head but reaches a TrackML score of 0.806 with the equal-token decode and a double-majority efficiency of 0. p. 2 In contrast to approaches based on explicit hit graphs or sequential track construction, the
method presented here combines sparse full-event parameter inference with a second learned local refinement
stage to resolve residual ambiguities in track assignment. HEPTv2 [26] reports a double-majority tracking efficiency of 98. 1 Introduction
Particle reconstruction at the Large Hadron Collider traditionally relies on a sequence of specialized algorithms
that transform low-level detector measurements into tracks, vertices, jets, flavour-tagging information, and
other reconstructed physics objects before event-level inference is performed. Modern machine-learning
approaches have begun to replace individual components of this chain, but tracking studies commonly focus on
a restricted subset of charged particles, often imposing transverse-momentum thresholds around pT ∼ 1 GeV. The question behind this work is whether a learned representation can instead be constructed from the full
detector event, retaining information from the complete charged-particle population rather than restricting
reconstruction to a high-pT subset. AlphaHits-TrackFormer therefore processes the full pixel and strip hit
cloud and attempts to reconstruct all reconstructable charged particles represented in the TrackML sample,
without imposing a pT threshold on the target tracks. The question behind this work is whether a learned representation can instead be constructed from the full
detector event, retaining information from the complete charged-particle population rather than restricting
reconstruction to a high-pT subset. AlphaHits-TrackFormer therefore processes the full pixel and strip hit
cloud and attempts to reconstruct all reconstructable charged particles represented in the TrackML sample,
without imposing a pT threshold on the target tracks.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 6 ---
Our earlier TrackFormer and HiggsFormer studies explored this idea from complementary directions TrackFormer investigated Transformer-based assignment of inner-tracker hits to charged-particle trajectories HiggsFormer showed that a related architecture could be retrained to classify Higgs events directly from raw inner-tracker hits bypassing explicit track and object reconstruction Trained on the raw inner-tracker hits of simulated ttH¯ (H → b¯b) and tt¯ events, it reached AUC of 0.855, comparable to what a Particle Transformer [5] on Delphes tracker objects reaches at the 40% b-tagging working point (0. p. 1 Trained on the raw
inner-tracker hits of simulated ttH¯ (H → b
¯b) and tt¯ events, it reached AUC of 0. 855, comparable to what a
Particle Transformer [5] on Delphes tracker objects reaches at the 40% b-tagging working point (0.Where the paper says it
--- Page 7 ---
AlphaHits-TrackFormer reconstructs tracks in complete TrackML events of up to 1.5 × 10 5 hits A learned Router sparsifies the attention a global Tracker model predicts binned and continuous track parameters for every hit a charge-aware preconditioner and a learned Gate select the ambiguous regions and a local Refiner resolves them AlphaHits-TrackFormer reaches TrackML scores of 0. p. 1 A learned Router sparsifies the attention, a global Tracker model
predicts binned and continuous track parameters for every hit, a charge-aware preconditioner and a
learned Gate select the ambiguous regions, and a local Refiner resolves them. Trained on 7 830 events,
it reaches TrackML scores of 0.Where the paper says it
--- Page 8 ---
The AlphaHits-TrackFormer reconstruction pipeline is hierarchical A learned Router first assigns detector hits to overlapping logical groups that determine the sparse attention structure A full-event Tracker then predicts track parameters for every hit and maps the hits into a discrete track-parameter space Finally, a learned Gate identifies regions that require additional processing and a compact local Refiner resolves the remaining combinatorial ambiguities
--- Page 9 ---
The Router is a small MLP that processes hits independently It receives the continuous hit features together with learned embeddings describing the detector volume layer and module and predicts membership in two of four × 4 × 32 groups made in hit-radius track-pseudorapidity and track-azimuth The Router therefore serves two purposes it makes full-event attention computationally tractable and introduces a learned locality structure determining which detector regions can communicate directly
--- Page 10 ---
Hits, used as an input are represented as short vectors of their position and local geometry They are embedded by one linear layer into d=256 vectors and passed through a stack of Transformer encoder layers The output of the encoder is one state per hit and every head of both models reads these per-hit states AlphaHits adds learned embeddings of the detector volume layer and module and eight cluster-shape features for every input hit
--- Page 11 ---
AlphaHits predicts, for every hit its particle’s kinematics, production-vertex displacement and charge, and its member jet’s axis pT and secondary-vertex flight distance next to the track token The auxiliary terms enter the loss with fixed weights outside the main objective
--- Page 12 ---
AlphaHits predicts for every hit its particle’s kinematics, production-vertex displacement and charge, and its member jet’s axis pT and secondary-vertex flight distance next to the track token The auxiliary terms enter the loss with fixed weights outside the main objective
--- Page 13 ---
AlphaHits is a permutation-equivariant Transformer architecture operating directly on detector hits building on the TrackFormer and HiggsFormer models The first HiggsFormer was an event classifier over raw hit coordinates (x, y, z) with a single ttH¯ (H → b¯b) vs tt¯ head plus an auxiliary HT regression head for the ‘big’ model
--- Page 14 ---
Each hit is the seven-dimensional vector: position in metres transverse radius azimuth as (cos φ,sin φ) and pseudorapidity The model sees positions only no detector identifiers cluster shape charge or timing
--- Page 15 ---
AlphaHits is a permutation-equivariant Transformer architecture operating directly on detector hits building on the TrackFormer and HiggsFormer models It is pretrained on tracking tasks only (Stage 1) and fine-tuned with nine heads to predict tracks, jets with flavour, HT, Emiss T, leptons and the event class in one forward pass (Stage 2)
--- Page 16 ---
The first HiggsFormer was an event classifier over raw hit coordinates (x, y, z) with a single ttH¯ (H → b¯b) vs tt¯ head plus an auxiliary HT regression head for the ‘big’ model AlphaHits builds on the tracking backbone of Section 4
--- Page 17 ---
All training of the AlphaHits-TrackFormer models was conducted on an NVIDIA H200 card and took roughly 1 week for the best model The Router is first pretrained for one epoch using only the supervised routing objective The Router and Tracker are subsequently trained jointly for 50 epochs using the AdamW optimizer with an initial learning rate of 3 × 10-4, weight decay of 1 × 10-4, and cosine learning-rate annealing
--- Page 22 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 24 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 26 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 27 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 28 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 29 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 30 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 31 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 32 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 33 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 34 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 35 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 36 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 37 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 38 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 39 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is trained with a multi-task objective combining discrete track-parameter prediction, hit classification, continuous track-parameter regression, charge prediction, and routing supervision
--- Page 40 ---
The Router is supervised through Lrouter a soft-target cross-entropy loss over the 512 routing groups The global stage is …
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
-
Scalable Full-Event Tracking for High Occupancy Events: The AlphaHits-TrackFormer architecture processes
all pixel and strip detector hits in complete events containing up to 1.5 x 105 hits under HL-LHC-like pileup conditions,
enabling reconstruction of all reconstructable charged particleswithout imposing a transverse-momentum selection on the target tracks.
-
Multi-Task Learning from Hit Representations: The system can perform simultaneous reconstruction and classification by using
a common representation of low-level detector information
to predicttracks, jets, jet flavour, and event-level observables
from a shared hit-level representation. -
Inclusive Low-pT Charged Particle Reconstruction: The model reconstructs tracks down to approximately 150 MeV in the supplied TrackML samples, including the
highly challenging low-pT population,
which otherwise constitutes over 90 percent of charged particles. -
Enhanced Event Classification Performance: By extending the architecture with auxiliary heads, AlphaHits achieves an AUC of 0.944 for Higgs event classification when including calorimeter towers as additional inputs, demonstrating performance
comparable to that of a Particle Transformer (ParT) baseline using fully reconstructed inputs.
-
Hierarchical Reconstruction Pipeline: The system employs a two-stage process: Stage A (Global) uses sparse attention to predict track parameters over all hits, and Stage B (Local) uses
a learned Gate
and alocal Refiner
to resolve residual ambiguities, specifically resolvingcombinatorial ambiguities required for track assembly.
-
Efficient Computation via Sparse Attention: The model utilizes a
learned Router
that assigns each hit to groups to make attention sparse, avoiding dense all-to-all self-attention costs, which is crucial when operating on events with up to 105 hits. -
Transferable Representation Learning: The approach establishes
transfer across physics objectives
by first training the backbone on tracking alone and then fine-tuning it jointly with nine heads for multi-task reconstruction and classification, creating ahit-level foundation model for collider physics.
Abstract
We introduce AlphaHits, a Transformer-based approach towards hit-level foundation models for LHC physics, building on the TrackFormer and HiggsFormer architectures. AlphaHits operates directly on detector hits and demonstrates scalable full-event charged-particle reconstruction and multi-task event reconstruction and classification from low-level detector information. AlphaHits-TrackFormer reconstructs the full charged-particle content of complete TrackML events without imposing a transverse-momentum selection on the target tracks. It processes both pixel and strip hits in events containing up to 1.5 times10 5 hits under HL-LHC-like pileup conditions and assigns hits to individual reconstructed trajectories across its kinematic range. This includes the challenging low- p T population down to approximately 150, MeV in TrackML samples. It achieves a TrackML score of 0.959 over all reconstructable tracks, representing, to our knowledge, the highest reported TrackML score for full-detector, inclusive charged-particle reconstruction. Building on the TrackFormer architecture, AlphaHits is pretrained on charged-particle tracking and fine-tuned for multi-task reconstruction and classification. From a shared hit-level representation, it predicts tracks, jets, jet flavour, and event-level observables. AlphaHits achieves an AUC of 0.944 for Higgs event classification, with performance comparable to a Particle Transformer (ParT) baseline using fully reconstructed inputs. It also achieves performance comparable to Delphes-based baselines for jet reconstruction, jet flavour classification, and missing transverse momentum regression. These results highlight the potential for scalable full-event tracking and multi-task learning directly from detector hits, representing a promising step towards transferable hit-level representations and general-purpose foundation models for LHC physics.
Sources
- Large Physics Models: Towards a collaborative approach with Large Language Models and Foundation Models
- TrackFormers: In Search of Transformer-Based Particle Tracking for the High-Luminosity LHC Era
- Hits to Higgs: Hit-Level Higgs Classification from Raw LHC Detector Data Using Higgsformer
- TrackFormers Part 2: Enhanced Transformer-Based Models for High-Energy Physics Track Reconstruction
- Prompting Particle Physics: Tokenized Multi-modal Foundation Models for Combinatorially Many Tasks
- Particle Transformer for Jet Tagging
- Masked Particle Modeling on Sets: Towards Self-Supervised High Energy Physics Foundation Models
- Finetuning Foundation Models for Joint Analysis Optimization
- Foundation model for particle collision data analysis
- OmniJet-$\alpha$: The first cross-task foundation model for particle physics
- Solving Key Challenges in Collider Physics with Foundation Models
- OmniLearned: A Foundation Model Framework for All Tasks Involving Jet Physics
- Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics
- Machine-learned particle flow as a foundation model for collider physics
- Performance of a Geometric Deep Learning Pipeline for HL-LHC Particle Tracking
- Towards a realistic track reconstruction algorithm based on graph neural networks for the HL-LHC
- Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph and image data
- Reconstruct everything everywhere all at once: An end-to-end model for future colliders
- High Pileup Particle Tracking with Object Condensation
- Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
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