Researchers recast particle track fitting as a language translation task, and a decoder-only transformer improved track-parameter accuracy by an order of magnitude on synthetic data while identifying multi-track events with 98% accuracy
Related research and updates
Synopsis
This exploratory study formulates particle track fitting as translating the language of detectors into the language of physics, designs a simple decoder-only transformer with a custom tokenizer, and reports that on synthetic datasets it outperformed a simple regression technique by an order of magnitude in track-parameter accuracy while being on par with RANSAC, identified multi-track events with 98% accuracy, and produced a zenith angle distribution for real cosmic-muon events in excellent agreement with expectations.
Figure 1: Schematic diagram of a Resistive Plate Chamber showing the strip placements on the top ( x x -side) and at the bottom ( y y -side). In the prototype detector, there are 32 strips on either planes. The strip numbers on theses planes encode the x x , y y coordinates of the particle passage through these detectors. 12 such RPCs are stacked on top of each other with a gap in between each one. The location of the RPCs along the height of the stack gives the z z -coordinate.
arXivInterpretation
The authors recast particle track fitting as a language translation problem, translating the language of detectors into the language of physics, and designed a simple decoder-only architecture with a custom tokenizer for this task. The traditional pipeline collects individual detector hits, recombines them into tracks, and fits them to a model function; this work rewrites that pipeline as a sequence-to-sequence translation task, borrowing the grammar-and-semantics capability of large language models. The paper positions itself as an exploratory study, describes the architecture and tokenizer design, and evaluates the model on synthetic datasets and real cosmic-muon data.
On synthetic datasets, the model outperformed a simple regression technique by an order of magnitude in track-parameter accuracy while being on par with the random sample consensus (RANSAC) method. This comparison provides a quantitative positioning of the formulation relative to a baseline regression method and a classical robust fitting method. The conclusion rests on track-parameter accuracy evaluation on synthetic datasets, described in the paper as an order of magnitude better and on par with RANSAC.
The model identified multi-track events with an accuracy of 98%, and on real cosmic-muon events the zenith angle distribution showed excellent agreement with expectations. Multi-track events appear natural to represent under this formulation, and the model transferred from synthetic data to real cosmic-muon data while remaining consistent with expectations. Multi-track accuracy is reported as 98%; the real-data test compares the zenith angle distribution of cosmic-muon events with expectations.
Perspective
The work targets track fitting in particle physics experiments, applicable to reconstructing track parameters from detector hits and handling multi-track events, with evaluation built on synthetic datasets and real cosmic-muon data. For researchers and engineers who want to bring sequence modeling into detector data analysis, the formulation offers a reusable idea: encode detector information as a sequence with a custom tokenizer and let a decoder-only transformer output physics parameters. The paper notes that further studies with complex geometries are required to understand the full potential of the technique.
As an exploratory study, the conclusions rest on synthetic datasets and real cosmic-muon data, and the paper itself notes that further studies with complex geometries are required to understand the full potential of the technique. Readers may watch how track-parameter accuracy and multi-track identification behave under more complex detector geometries and higher multiplicity; how much the design choices of the custom tokenizer and decoder-only architecture affect the results; and whether the agreement of the zenith angle distribution in the real-data test also holds for other physics quantities.