acceptodds
Under review as a conference paper at ICLR 2027

ML-QTrack: Learning-Guided Global Optimization for Full-Detector Particle Tracking

Abstract

Charged-particle track reconstruction is a fundamental yet challenging problem in high-energy physics, with full-detector tracking posing a particularly difficult trade-off between reconstruction quality and computational cost. Learning-based methods can efficiently process large candidate graphs but may produce substantial fake trajectories, while global optimization enforce stronger trajectory-level consistency at a higher computational cost. We propose ML-QTrack, a learning-guided global optimization framework that combines graph purification with trajectory-level QUBO optimization. On the full-event, full-detector TrackML benchmark, ML-QTrack achieves 98.1% efficiency and reduces the fake rate from 36.71% to 6.87% compared with an HGNN-based method using the same candidate graph, while completing reconstruction in per event. These results show that learning-guided hypothesis reduction can make trajectory-level global optimization computationally practical for large-scale full-detector tracking, while substantially improving reconstruction quality.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning, Quantum Inspired Algorithm, Quadratic Unconstrained Binary Optimization, Track Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_71G9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ML-QTrack, a learning-guided global optimization framework for charged-particle track reconstruction in high-energy physics. The method combines metric-learning-based candidate graph construction, LightGBM-based graph purification, and trajectory-level Quadratic Unconstrained Binary Optimization. A post-optimization refinement stage further resolves duplicate and fragmented trajectories. Experiments on the TrackML benchmark show that ML-QTrack achieves 98.1% reconstruction efficiency and a 6.87% fake rate in the full-event, full-detector setting, substantially reducing the fake rate compared with an HGNN-based baseline. The paper also provides component ablations, comparisons across different detector configurations, and runtime analyses demonstrating second-scale event reconstruction.

Critical strengths

S1. The combination of learning-based graph purification and global optimization is well motivated. ML-QTrack addresses the complementary limitations of learning-based tracking and combinatorial optimization. Using LightGBM to reduce the candidate hypothesis space before QUBO optimization provides a practical approach to improving computational efficiency while preserving reconstruction quality.

S2. The experimental results are encouraging under realistic full-detector conditions. ML-QTrack achieves 98.1% reconstruction efficiency and reduces the fake rate from 36.71% to 6.87% compared with HGNN in Table 4. The results are particularly interesting because the method operates on full events covering both barrel and endcap detector regions, rather than relying only on simplified detector configurations.

Critical weaknesses

W1. The contribution of the trajectory-level refinement stage is not sufficiently examined. Section 3.3 introduces a post-optimization refinement procedure to resolve duplicate trajectories and reconnect fragmented segments. However, the ablation in Table 1 mainly focuses on graph purification and global optimization, without separately evaluating the refinement stage. Since duplicate suppression and fragment recovery can directly affect reconstruction efficiency and fake rate, it is unclear how much of the final improvement comes from QUBO optimization versus this additional post-processing procedure. An ablation comparing reconstruction results before and after trajectory refinement would help clarify their individual contributions.

W2. The solution quality of the QUBO optimization is not sufficiently evaluated. Section 3.2 employs simulated bifurcation to solve the trajectory-level QUBO problem, while Appendix D specifies a fixed configuration of 300 iterations and 10 parallel trajectories. However, the paper does not examine how closely the resulting solutions approximate the QUBO optimum or how reconstruction performance changes with the solver's computational budget. Although Table 1 demonstrates improvements over greedy selection, it remains unclear whether the reported performance is limited by the QUBO formulation or the heuristic solver. An analysis of solution quality and convergence under different solver budgets would help clarify the effectiveness and reliability of the proposed global optimization approach.

Questions for authors

Q1. What is the individual contribution of the trajectory-level refinement stage? Could the authors report reconstruction efficiency and fake rate before and after duplicate merging and fragment recovery?

Q2. How sensitive are the reconstruction results to the simulated bifurcation solver's iteration count and number of parallel trajectories? Could the authors provide convergence analyses or compare solution quality against a stronger solver on smaller QUBO instances?

1 comment
AuthorsOpenReview ↗

Quantifying Trajectory Refinement Contributions and bSB Solution Quality

We thank the reviewer for recognizing our learning-guided QUBO approach. We have added controlled refinement ablations and solver-quality experiments to address both concerns.

Q1. Contribution of trajectory refinement

Revision (§§3.3 and 4.3, Table 1). We compare reconstruction before and after refinement, keeping purification and QUBO selection fixed. Truth-assisted duplicate-hit removal serves as a diagnostic control.

Input (purified, QUBO selection)RefinementEfficiency (%)Fake rate (%)
Full inputOff83.4030.28
Full inputOn97.716.11
Truth-assisted deduplicationOff98.121.45
Truth-assisted deduplicationOn98.411.14

On full inputs, refinement improves efficiency by 14.31 percentage points and reduces the fake rate by 24.17 points, while gains after deduplication are much smaller (+0.29 / −0.31 points). This indicates that same-layer hit multiplicity contributes substantially to the ambiguities addressed by refinement.

QUBO globally selects inter-layer triplet hypotheses, whereas refinement resolves duplicate and fragmented trajectories not directly modeled by QUBO. Under identical refinement settings, QUBO outperforms Greedy both before (83.40% / 30.28% vs. 74.86% / 37.12%) and after refinement (97.71% / 6.11% vs. 94.24% / 7.53%), demonstrating complementary contributions. Appendix H details the recovery criteria; the ablation measures the combined refinement stage.

Q2. Solver-budget sensitivity and solution quality

Revision (Appendix F). We added budget sensitivity experiments on 20 fixed event-level QUBOs and exact-optimum comparisons on 400 connected components. Both use truth-assisted deduplication without refinement to isolate QUBO optimization.

Solver-budget sensitivity. Candidate graphs, QUBO coefficients, decoding, and evaluation remain fixed.

Solver configurationEfficiency (%)Fake rate (%)Mean QUBO energy
bSB, 100 iterations, batch 1097.642.05−6260.6
bSB, 300 iterations, batch 1098.121.45−6846.2
bSB, 600 iterations, batch 1098.121.39−6882.6
bSB, 1000 iterations, batch 1098.161.36−6891.4
bSB, 300 iterations, batch 198.181.36−6837.3
bSB, 300 iterations, batch 4098.111.41−6850.8
neal, 10 reads98.161.82−6713.8

Increasing iterations from 100 to 300 substantially improves energy and reconstruction quality. From 300 to 1000, mean energy decreases by another 45.2, but efficiency improves by only 0.04 points and fake rate by 0.09 points. Batch sizes yield comparable tracking performance. Thus, 300 iterations offer a practical quality–budget trade-off, although further optimization remains possible.

The neal control achieves slightly higher efficiency but worse fake rate and QUBO energy than bSB300. Its computational budget is not matched, so this is an alternative-solver reference rather than a claim of universal superiority.

Exact-optimum comparison. We exhaustively enumerated 400 complete QUBO connected components, selecting the 20 largest eligible components (at most 20 variables) per event. Complete components preserve all couplings.

Solver configurationExact optima found
bSB, 300 iterations, batch 10286/400 (71.5%)
bSB, 1000 iterations, batch 10361/400 (90.3%)
neal, 10 reads224/400 (56.0%)

Increasing the bSB budget raises exact-optimum success from 71.5% to 90.3%, compared with 56.0% for neal. These results establish solution quality on tractable components, not global optimality for full-event QUBOs.

Together, the experiments support bSB with 300 iterations and 10 parallel trajectories as a practical configuration, while showing that larger budgets improve optimization quality with diminishing reconstruction gains.

Reviewer_ReHgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a learning-based quadratic unconstrained binary optimization approach as a middle ground solution to full-detector tracking and shows good empirical performance on the CPU-scale, against a HGNN baseline using the TrackML challenge. They positioning is that learning based methods can generate a large number of tracks but also generate fake tracks at a higher rate and traditional global optimization produces strong physics-performance for the reconstructed tracks but are computationally heavy; so the proposed method aims to combine the best of both and find a good trade-off/middle ground.

Critical strengths

The ML novelty is minimal while the overall idea from a systems engineering perspective is good. The learning graph is directly borrowed from GNN4ITK, and the purification is simply shallow LightGBM models, and while the combination to QUBO is a good trade-off the QUBO hamiltonian is blind to the learned LightGBM probabilities and this seems like a missed opportunity. So overall novelty for the AI community outside HEP is low.

Missing evaluations depth: The results are at the empirical proof-of-concept level. There are not enough competitive baselines (eg. GNN4ITK, Interaction Networks etc.) apart from HGNN, there are not enough ablations that clearly quantify the contribution of QUBO.

Critical weaknesses

good idea, clean presentation, but empirical results, not enough baselines and minimal novelty for an AI conference.

there are concerns about missing baselines, missing ablations, and this being an empirical study rather than a full-scale tracking alternative;

Questions for authors

Please expand the results to more relevant baselines as mentioned in the weaknesses section.

Please add ablations results to show the individual contributions of each aspect of your algorithm.

The learned LightGBM potentials are currently discarded at the QUBO term stage. Have you tried using the learned edge/triplet scores directly (or a suitable projection) of it when constructing the QUBO potentials?

Please check or clarify the fake rate discrepancies between tab 1, tab 4 and fig 10.

What happens with the current setup for sub 1GeV tracks given the pT>1GeV choice for the training labels?

Consider moving the AI Disclosure statement after the main text and before the references as per the official guidelines instead of at the end of the abstract.

1 comment
AuthorsOpenReview ↗

Strengthening Baselines, Component Ablations, and Learned QUBO Potentials

We thank the reviewer for the careful assessment and valuable suggestions. We have strengthened the reconstruction baselines, expanded the component ablations, and investigated learned QUBO potentials. The methodological contribution of ML-QTrack lies in integrating learned hypothesis reduction with trajectory-level global optimization, rather than introducing a new classifier architecture. At 100,000 hits, LightGBM reduces the QUBO from approximately 2.53 million to 29,700 variables while retaining 98.16% of true candidate edges (§4.2), enabling tractable full-detector reconstruction.

Q1. Additional reconstruction baselines

Revision (§4.4 and Appendix L). We reproduced BC-HGNN-GMM using its released TrackML-1GeV configuration and evaluated both methods on 20 common events:

MethodEfficiency (%)Fake rate (%)
ML-QTrack, with refinement97.941.40
HGNN, candidate graph94.403.01
HGNN, purified graph86.555.67

We distinguish this common-event evaluation from literature-reference comparisons with prior QUBO methods, Object Condensation, Transformer tracking, and published HGNN results (§4.4, Tables 2–4). The latter differ in detector coverage, event size, or evaluation requirements and are not presented as fully matched benchmarks. The common-event results favor the complete ML-QTrack pipeline, but do not isolate QUBO selection from HGNN-specific post-processing differences.

Q2. Individual component contributions

Revision (§4.3, Table 1). We expanded the ablation to evaluate graph purification, QUBO selection, and combined trajectory refinement, using truth-assisted duplicate-hit removal as a diagnostic control.

Without purification or refinement, QUBO improves efficiency from 89.54% to 93.50% and reduces the fake rate from 24.81% to 18.75% relative to Greedy on deduplicated inputs. Adding purification improves QUBO performance to 98.12% efficiency / 1.45% fake rate.

On full inputs, QUBO outperforms Greedy both without refinement (83.40% / 30.28% versus 74.86% / 37.12%) and with identical refinement (97.71% / 6.11% versus 94.24% / 7.53%). These controlled comparisons quantify the contributions of graph purification, QUBO-based selection, and the combined geometric refinement stage.

Q3. Using LightGBM scores in the QUBO objective

Revision (§3.2 and Appendix G). The primary QUBO uses geometric potentials, while LightGBM provides learned candidate filtering. To investigate tighter integration, we project the probabilities of a triplet's two constituent edges into its unary coefficient:

$ a_i'=a_i-\frac{\lambda}{2} \left[\ell(p_{d_1})+\ell(p_{d_2})\right], $

where ℓ(p)\ell(p) denotes normalized, clipped log-odds.

On 20 truth-deduplicated events without refinement, the original geometric formulation (λ=0\lambda=0) achieves 98.12% efficiency / 1.45% fake rate, compared with 98.16% / 1.35% at λ=0.10\lambda=0.10 and 98.31% / 1.13% at λ=0.25\lambda=0.25. All other reconstruction settings remain fixed.

These results suggest that learned confidence can complement geometric QUBO potentials. The original model represents geometry-driven global optimization, whereas the extension combines geometric consistency with learned edge confidence. We retain the original formulation as the primary configuration because selecting and validating a default λ\lambda requires an independent validation set and subsequent testing.

Q4. Fake-rate differences across tables and figures

Revision (§§4.1, 4.3–4.4 and Appendix J). We clarified the evaluation populations and processing settings. Table 1 reports a 1.45% event-level fake rate with truth-assisted duplicate-hit removal and no refinement, versus 6.11% on full inputs with refinement, consistent with Table 4.

Figure 10 instead reports a restricted distribution binned by truth-particle pTp_T, using deduplicated inputs without refinement. It excludes tracks associated with sub-GeV particles and tracks without an assignable truth pTp_T. These binned fractions are therefore not directly comparable to the complete event-level fake rate.

Q5. Tracks below 1 GeV

Clarification (Appendix A.1, §4.1, Appendix J, and Conclusion). Our evaluation targets particles with pT≥1p_T\geq1 GeV. Although lower-pTp_T hits remain in full-input events, LightGBM purification and the geometric QUBO formulation are designed for this momentum range. We acknowledge that extending reconstruction below 1 GeV may require adapting candidate filtering and geometric compatibility measures to accommodate greater transverse curvature, and identify this as future work.

Q6. AI Disclosure

Revision. We moved the AI Disclosure to a standalone section between the conclusion and references. It clarifies the use of AI-assisted tools in manuscript preparation, with the authors responsible for verifying all scientific content and results.

Reviewer_JTSpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a particle-tracking pipeline that combines learned candidate filtering with combinatorial optimization. It starts from a metric-learning candidate graph and uses LightGBM to remove unlikely hit connections. It then constructs overlapping three-hit trajectory segments and selects among them through a quadratic unconstrained binary optimization (QUBO) objective. A simulated-bifurcation solver approximately optimizes this objective, and geometric procedures then merge duplicate tracks and reconnect fragments.

The main contribution is making triplet-based optimization computationally manageable on full-detector TrackML events. The paper reports 98.1% reconstruction efficiency, a 6.87% fake rate, and approximately 1.68 seconds per event. The evaluation includes 100 held-out events, component ablations, and comparisons under several detector configurations.

Critical strengths
  1. The hypothesis reduction is substantial and directly relevant to computational feasibility. The paper reports reducing candidate edges from approximately 436,300 to 26,400 while retaining 98.16% of true edges. At 100,000 hits, the number of QUBO variables decreases from approximately 2.53 million to 29,700.

This is a useful systems result: an inexpensive classifier removes connections that would otherwise generate many combinations of candidate segments. It explains why the downstream optimization becomes practical.

  1. The paper moves beyond restricted detector configurations and reports useful event-level variability. Evaluating full events across barrel and endcap regions is more informative than demonstrating the method only on small detector sectors. The standard deviations reported across 100 test events also strengthen the empirical evidence.

The attempt to handle duplicate measurements and fragmented tracks without using particle identities during recovery is practically relevant, although the contribution of that recovery stage needs to be isolated.

Critical weaknesses
  1. The headline comparison does not establish that global optimization causes the reported fake-rate reduction.

The paper emphasizes reducing HGNN’s fake rate from 36.71% to 6.87%. The underlying HGNN report states, however, that its reported fake rates are intended for model-to-model comparisons rather than final track-building performance, and that track cleaning and ambiguity resolution can greatly reduce them. ML-QTrack includes a dedicated duplicate-merging and fragment-recovery stage. Comparing these numbers without equivalent cleanup does not isolate the benefit of QUBO optimization.

Using the same initial candidate graph is helpful, but it does not control for differences in filtering, track assembly, and refinement. A matched comparison should evaluate both methods with comparable cleanup and show ML-QTrack before and after refinement.

The internal greedy comparison also needs clarification. Appendix G says that track growth selects the continuation with the smallest quadruplet strength, whereas Equation (4) defines larger compatibility scores as better. If “strength” means the positive compatibility score, the baseline appears to choose the less compatible continuation. If it means the negative QUBO coefficient, the procedure may be appropriate, but that definition must be explicit.

These issues are central because the paper attributes its improvement to global trajectory optimization. The current comparisons do not establish that attribution.

  1. The stated conflict rule penalizes segments that can belong to the same valid trajectory.

Section 3.2 defines two triplets as conflicting when they share at least one hit but do not form a valid quadruplet. Consider a valid five-hit trajectory with consecutive triplets T1=(h1,h2,h3),T2=(h2,h3,h4),T3=(h3,h4,h5)T_1=(h_1,h_2,h_3),\qquad T_2=(h_2,h_3,h_4),\qquad T_3=(h_3,h_4,h_5). The adjacent pairs are compatible continuations. The first and third triplets, however, share one hit and do not form a four-hit continuation. Under the written definition, they therefore receive a positive conflict penalty despite belonging to the same trajectory.

This case does not depend on noisy measurements: it occurs in an ordinary chain of consecutive triplets. With zero unary penalties and unit compatibility rewards, selecting all three has energy H(1,1,1)=ζ−2H(1,1,1)=\zeta-2, while selecting only the first two has energy H(1,1,0)=−1H(1,1,0)=-1. When the conflict penalty exceeds one, the objective therefore prefers dropping a valid segment in this example. Smaller penalties do not remove the underlying issue that legitimate overlap is treated as conflict.

The manuscript does not specify the numerical conflict penalty or explain how the implementation distinguishes compatible one-hit overlap from branching conflicts. If the implementation uses a narrower rule, the formal definition needs correction. Otherwise, the claim that the objective enforces trajectory consistency requires reconsideration, including whether post-processing repairs fragmentation introduced by the objective itself.

Questions for authors
  1. Baseline comparability: Was HGNN rerun on the same test events, or were its published numbers reused? What happens when HGNN receives comparable track cleaning and ambiguity resolution? Please report ML-QTrack before and after trajectory refinement under the headline full-detector protocol.
  2. Conflict construction: How does the implementation handle the five-hit example above? Please provide the exact conflict predicate and the value of the conflict penalty, and state whether that penalty was selected on validation events.
  3. Greedy ablation: Does “smallest quadruplet strength” mean minimizing the negative QUBO coupling or minimizing the positive compatibility score? Please clarify the implementation and report the corrected or verified greedy comparison with identical refinement and evaluation settings.
  4. Input features: Appendix A.1 says spatial coordinates are retrieved from the truth table. Are the classifier and geometric optimization supplied with measured hit coordinates or with simulated true intersection coordinates? TrackML distinguishes these quantities, so the exact columns used during training and test inference matter. I am asking for clarification, not asserting that leakage occurred.
  5. Evaluation and timing: Which protocols produce the fake rates in Table 1, Table 4, and Appendices H–I? Please also confirm whether the reported total runtime includes trajectory refinement and whether metric-learning graph generation was timed on the same CPU configuration.
3 comments
AuthorsOpenReview ↗

Clarifying QUBO Formulation and Strengthening Fair Comparisons

Thank you for your careful review. We have revised the manuscript to clarify the QUBO formulation, strengthen the controlled comparisons, and distinguish the evaluation and runtime settings.

Q1. HGNN comparability and trajectory refinement

We agree that the published HGNN comparison does not isolate QUBO's contribution. HGNN's reported 36.71% fake rate reflects model-level performance that may improve with additional track cleaning. We now explicitly distinguish published results from matched evaluations.

On 20 common events (Appendix L), ML-QTrack achieves 97.94% efficiency / 1.40% fake rate, compared with 94.40% / 3.01% for HGNN. Applying LightGBM purification to HGNN yields 86.55% / 5.67%, potentially due to disrupted graph connectivity. This comparison supports the performance of our complete pipeline, rather than an isolated QUBO advantage over HGNN.

Our expanded ablation (§4.3, Table 1) distinguishes the two stages: QUBO globally selects inter-layer triplet hypotheses, while refinement resolves duplicate and fragmented trajectories associated with same-layer hit multiplicity, which the QUBO does not directly model. Refinement improves full-input performance from 83.40% / 30.28% to 97.71% / 6.11%, but yields a much smaller gain after truth-assisted deduplication (98.12% / 1.45% to 98.41% / 1.14%). The controlled Greedy comparison in Q3 separately establishes the contribution of QUBO selection.

Q2. Conflict construction and the five-hit example

We agree that the original conflict description was overly broad. However, the submitted Figure 3 and actual implementation already classified end-to-start overlaps as unrelated. The implementation already distinguishes end-to-start overlaps from branching conflicts, but the original textual definition did not adequately describe this distinction. We have therefore revised the formal conflict predicate to make the mathematical formulation consistent with the implementation.

Revision (§3.2 and Appendix C): We clarify that compatible triplets receive bij=−Sijb_{ij}=-S_{ij}, conflicting triplets receive bij=ζb_{ij}=\zeta, and unrelated pairs receive bij=0b_{ij}=0. In the reviewer's example, T1=(h1,h2,h3)T_1=(h_1,h_2,h_3) and T3=(h3,h4,h5)T_3=(h_3,h_4,h_5) are unrelated, so b13=0b_{13}=0. With unit continuation rewards and zero unary terms, H(1,1,1)=−2<H(1,1,0)=−1H(1,1,1)=-2 < H(1,1,0)=-1, correctly favoring the five-hit trajectory.

Q3. Greedy baseline and controlled ablation

We corrected the ambiguous phrase “smallest quadruplet strength” in Appendix I. Greedy growth selects the continuation with the largest compatibility score SijS_{ij}, equivalently the most negative coupling bij=−Sijb_{ij}=-S_{ij}, using triplet quality to break ties.

We also added a controlled Greedy-versus-QUBO ablation (§4.3, Table 1). Both methods use identical candidate triplets, unary scores, geometric compatibility information, and refinement procedures, differing only in their selection strategy.

SelectionRefinementEfficiency (%)Fake rate (%)
GreedyOff74.8637.12
QUBOOff83.4030.28
GreedyOn94.247.53
QUBOOn97.716.11

With identical refinement, QUBO improves efficiency by 3.47 percentage points and reduces the fake rate by 1.42 points. QUBO also outperforms Greedy without refinement, demonstrating the benefit of global selection independently of the recovery stage.

Q4. Source of classifier and geometric inputs

We clarified the coordinate sources in Appendices A.1–A.2. Training and inference use measured hit coordinates (x,y,z)(x,y,z) from hits.csv, together with detector-volume information. Simulated intersection coordinates (tx,ty,tz)(tx,ty,tz) from truth.csv are never used as model or geometric reconstruction inputs. Truth information is used for supervision and evaluation, and explicitly for hit selection only in designated diagnostic experiments.

Q5. Fake-rate definitions and runtime reporting

We clarified the strict double-majority matching criterion (§4.1), under which a track must recover more than half of a particle's hits and contain more than half of its hits from that particle. The event-level fake rate is the fraction of unmatched reconstructed tracks.

The lower fake rates in Appendix J arise from a restricted truth-pTp_T-binned analysis using truth-assisted duplicate-hit removal without refinement. Tracks associated with sub-GeV particles or lacking an assignable truth pTp_T are excluded, so these values are not directly comparable to the event-level fake rate.

We also corrected the runtime reporting (§4.5): 1.68 s is the measured CPU subtotal for purification, QUBO construction, and bSB solving. Including approximately 0.5 s each for GPU graph generation and CPU refinement gives an estimated full-pipeline runtime of 2.68 s per event.

Thank you for the clarifications and the additional experiments. Your response helped me better understand the scope of the paper’s claims, and I will maintain my original rating.

AuthorsOpenReview ↗

Further Clarification of the QUBO Conflict Predicate and Penalty Selection

Thank you for considering our clarifications and additional experiments. We understand that these revisions have not changed your overall assessment. We would nevertheless like to highlight one further clarification in the revised manuscript: we have explicitly formalized the conflict predicate, including the distinction between branching conflicts and end-to-start overlaps, and specified the conflict penalty and its selection procedure. This makes the treatment of the five-hit example mathematically explicit and consistent with the implementation. We hope this clarification will also assist the Area Chair in assessing the technical soundness of the formulation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Evidence on ML-QTrack's Core Contributions: Controlled Ablations and Optimization Analysis

We sincerely thank all reviewers for their thoughtful assessments and constructive suggestions. In response, we have revised the manuscript and conducted additional experiments to strengthen the methodological analysis, clarify the contributions of individual components, and evaluate the reliability of the optimization procedure.

1. Core Contribution: Learning-Guided Global Optimization

The central contribution of ML-QTrack is a learning-guided hypothesis reduction strategy that makes triplet-level global optimization computationally practical for full-event, full-detector tracking. At 100,000 hits, LightGBM reduces the QUBO from approximately 2.53 million to 29,700 variables, while retaining 98.16% of true candidate edges (§4.2).

By suppressing unlikely hypotheses before optimization, the proposed framework enables QUBO to perform trajectory-level global selection using geometric compatibility and conflict constraints.

2. Distinguishing Global Selection and Refinement

We expanded the controlled ablations (§4.3, Table 1) to quantify the contributions of purification, QUBO selection, and trajectory refinement.

Under identical candidate and refinement settings, QUBO outperforms Greedy both without refinement (83.40% / 30.28% versus 74.86% / 37.12%) and with refinement (97.71% / 6.11% versus 94.24% / 7.53%). With identical refinement, QUBO improves efficiency by 3.47 percentage points and reduces the fake rate by 1.42 points.

The two stages address different challenges: QUBO globally selects inter-layer triplet hypotheses, whereas refinement resolves duplicate and fragmented trajectories associated with same-layer hit multiplicity. On full inputs, refinement improves performance of QUBO reconstruction from 83.40% / 30.28% to 97.71% / 6.11%. With truth-assisted duplicate-hit removal, the corresponding gain is much smaller (from 98.12% / 1.45% to 98.41% / 1.14%).

These controlled experiments demonstrate the complementary contributions of global selection and trajectory refinement, rather than attributing the overall improvement to a single component.

3. Solver Quality and Computational Budget

We added solver-budget sensitivity experiments on 20 fixed event-level QUBOs and exact-optimum comparisons on 400 tractable connected components (Appendix F).

Increasing bSB from 100 to 300 iterations improves efficiency from 97.64% to 98.12% and reduces the fake rate from 2.05% to 1.45%. Increasing the budget further to 1,000 iterations yields only marginal additional reconstruction gains (98.16% efficiency / 1.36% fake rate). Batch sizes of 1, 10, and 40 also produce comparable tracking metrics.

For the exact-optimum analysis, we selected 400 complete connected components containing at most 20 variables and obtained their exact minimum energies by exhaustive enumeration. bSB reaches these optima in **286 cases (71.5%) ** at 300 iterations and **361 cases (90.3%) ** at 1,000 iterations. The comparison with neal (224/400, 56.0%, using 10 reads) is provided as an additional solver reference rather than an equal-computational-budget comparison.

These results support 300 iterations as a practical computational budget. We do not claim global optimality for the full-event QUBO instances.

4. Additional Validation and Clarifications

We added a common-event comparison with HGNN using its released TrackML-1GeV configuration. On 20 common events, ML-QTrack achieves 97.94% efficiency / 1.40% fake rate, compared with 94.40% / 3.01% for HGNN. We clarify that this is a pipeline-level comparison and does not reproduce HGNN-specific post-processing; the published HGNN results under other protocols are therefore treated as literature references rather than fully matched benchmarks.

We also explored incorporating LightGBM confidence into QUBO unary potentials, improving performance from 98.12% / 1.45% to 98.31% / 1.13% in a controlled diagnostic study. We retain the original geometric formulation as the primary configuration pending further independent validation of the learned-score weight.

Finally, we clarified the conflict definition, Greedy baseline, fake-rate protocols, runtime reporting, and the current pT≥1p_T\geq1 GeV scope.

Overall, these revisions provide stronger evidence that learned hypothesis reduction, global QUBO selection, and geometric refinement play distinct but complementary roles, establishing ML-QTrack as a practical learning-guided framework for full-detector trajectory reconstruction

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.