Pith. sign in

REVIEW 5 major objections 8 minor 54 references

Physics losses only help calorimeter diffusion models when denoising stays the primary gradient; peer multi-task rules destroy fidelity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 02:19 UTC pith:BTO2ZWHM

load-bearing objection Solid controlled result that peer multi-task rules wreck shower fidelity while denoising-anchored blending does not; the “improves FPD and CFD” half is thinner than the abstract sells. the 5 major comments →

arxiv 2607.25060 v1 pith:BTO2ZWHM submitted 2026-07-27 cs.LG cs.AI

Lantern: Conflict-Aware Gradient Blending for Physics-Guided Diffusion Models in Calorimeter Simulation

Farzana Yasmin Ahmad , Vanamala Venkataswamy , Geoffrey Fox This is my paper
classification cs.LG cs.AI
keywords diffusion modelscalorimeter simulationphysics-informed learningmulti-task optimizationgradient blendinghigh-energy physicscorrelation metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Calorimeter shower simulation is a major bottleneck for the High-Luminosity LHC. Diffusion models can generate showers quickly, but their denoising loss is purely statistical, so a model can look good on that loss while getting the physics wrong. Standard physics-informed methods assume a closed-form law or hard per-sample constraint that a stochastic shower does not supply, and standard metrics miss correlation structure across layers and voxels. This paper introduces CFD, a single score for correlation fidelity, plus two soft physics losses (a variance-stabilized voxel residual and a graph Laplacian over detector geometry). It combines them with denoising via GradBlend, which keeps step size anchored to the denoising gradient while letting the physics term only steer direction. On CaloChallenge Dataset 2, peer multi-task rules inflate the main fidelity metric by factors of 2–100, whereas GradBlend admits the same signal without collapse and, with the Laplacian loss, improves both fidelity and correlation scores.

Core claim

On CaloChallenge Dataset 2, the same physics auxiliary losses that destroy shower fidelity when combined by task-symmetric multi-task rules (PCGrad, GradNorm, IMTL-G, ConFIG—FPD inflated 2–100×) can be admitted without regression by GradBlend, which anchors update magnitude to the denoising gradient and only lets the auxiliary steer direction; with the graph Laplacian loss, Lantern improves both FPD and voxel-wise CFD over the denoising-only baseline.

What carries the argument

GradBlend: a denoising-anchored gradient blend that normalizes both gradients to unit length, gates the auxiliary’s directional share by the conflict angle (full weight ≤120°, decay to 150°, off above), and sets step magnitude from the denoising norm alone times a conflict-dependent factor, so the auxiliary cannot inflate the learning rate or displace the generative objective.

Load-bearing premise

The hand-chosen conflict-angle gates, magnitude floor, and terminal denoising-only schedule that protect the primary objective will transfer beyond this one detector geometry, architecture, and observed conflict pattern.

What would settle it

Train the same two-stage DDPM+ViT setup with GradBlend and the Laplacian loss on CaloChallenge Dataset 3 or a hadronic shower geometry; if FPD and CFD still improve over denoising-only while peer multi-task rules still collapse, the claim holds; if GradBlend regresses or needs retuned gates, the primary-protection rule is overfit to Dataset 2.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Physics guidance for stochastic shower generators should treat denoising as primary and only admit auxiliaries through a magnitude-anchored directional blend, not peer multi-task rules.
  • CFD becomes a practical single-number check for layer- and voxel-wise correlation fidelity that FPD and KPD miss.
  • Conflicting auxiliaries (voxel residual) need a terminal denoising-only phase; non-conflicting ones (Laplacian) can stay on without schedule tuning.
  • Because guidance is training-only, fidelity gains add no generation-time cost and compose with faster samplers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Any generative setting where one statistical objective must stay primary and physics enters only as soft ensemble structure—not a hard PDE—may need the same magnitude-vs-direction split rather than scalar loss weights.
  • Measuring conflict angle over training could replace fixed angular thresholds with an adaptive gate tied to the observed geometry of the two gradients.
  • Joint training of voxel residual and Laplacian together remains untested; their interaction could either cancel residual conflict or create a new one.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 8 minor

Summary. The paper addresses physics guidance for diffusion-based calorimeter shower surrogates, where no per-sample PDE or conservation law exists to supervise training. It makes four contributions: (i) CFD, a normalized Frobenius distance between generated and reference voxel/layer correlation tensors; (ii) two auxiliary losses (a variance-stabilized voxel residual loss, Eq. 7, and a graph-Laplacian mismatch loss, Eq. 9); (iii) GradBlend, a denoising-anchored gradient combination rule that blends unit directions, anchors step magnitude to the denoising gradient norm, and gates the auxiliary's directional share by the conflict angle (Algorithm 1); (iv) the empirical finding that on CaloChallenge Dataset 2, task-symmetric multi-task rules (PCGrad, GradNorm, IMTL-G, ConFIG) inflate FPD by 2–100× while GradBlend admits the same auxiliary signal without regression, and — with the Laplacian loss under schedule S3 — improves both FPD and CFD over the denoising-only Base. A schedule ablation shows the conflicting voxel loss requires a terminal denoising-only phase while the aligned Laplacian loss is schedule-insensitive.

Significance. If the central comparative claim holds, this is a useful and honest contribution: the controlled design (identical architecture, schedule, losses, optimizer; only the combination rule varies) isolates the mechanism, the peer-method failures are large (2–100× FPD inflation) and mechanistically corroborated by the gradient diagnostics of Table 3, the PRDC breakdown in Table 5, and the loss-weighting baselines in Appendix D.2. The observation that general-purpose MTL rules mode-collapse when handed a physics auxiliary in the conflicting regime is a real, transferable warning for the fast-calorimeter-simulation community. Strengths to name: evaluation is against held-out Geant4 and public CaloChallenge submissions with the unmodified official pipeline; CFD is computed with one procedure across all models including prior surrogates; three seeds with mean±std are reported throughout; code is publicly released. The Base model is a strong baseline (FPD 30.87 vs CaloDREAM's 24.65), so the no-regression result is not a soft target. The weaker half — that Lantern *improves* over Base — is currently under-supported (major comments below), but the "admits physics signal without harm" claim, whi

major comments (5)
  1. [Abstract; §5.3, Table 1] The claim that Lantern (Laplacian, S3) "improves both FPD and CFD" is not supported by the paper's own numbers on the CFD side. CFD_vox is 0.120±0.005 vs Base 0.122±0.004: a 0.002 gap smaller than either seed std, and — more tellingly — smaller than Lantern's own variation across schedules in Table 2 (0.118/0.129/0.120 for S1/S2/S3), an internal non-seed variation roughly an order of magnitude larger than the claimed improvement. The honest reading of the CFD result is "no degradation," which is already a meaningful and headline-worthy result given the peer methods' failures. The abstract and §6 should be reframed accordingly (e.g., "matches Base on CFD while..."), or the authors must provide evidence that the CFD difference is real (more seeds, or a paired/per-seed analysis). As written, the strongest version of the claim could collapse to "matches Base" with more seeds, changing the pa
  2. [§5.3, Table 1; §7] The FPD improvement (30.87±1.26 → 25.97±1.44) rests on three shape-network seeds with a fixed energy network. A Welch t-test on the reported moments gives t≈4.4, p≈0.01, but with n=3 per arm the normality assumption is untestable and one additional seed could move this materially; §7 itself concedes three seeds limit power to resolve small differences, and a 16% FPD gap is exactly the small kind. Given that this improvement is half of the abstract's empirical claim, the paper needs either (a) more seeds (5–8) for the Base and Lantern-Laplacian rows, or (b) explicit downgrading of the claim to a trend consistent with the schedule-ablation mean (23.5–26.0 across S1–S3, Table 2, which does independently suggest Lantern-Laplacian sits below Base). Option (b) costs nothing and may suffice; option (a) is preferred.
  3. [§4.2, Algorithm 1; §4.3; Table 3] GradBlend's key thresholds look tuned to the observed conflict geometry of this task. Table 3 reports θ_conf ≈ 116–127° across methods, i.e., the measured conflict sits exactly inside the 120–150° linear-decay gate. The paper states the S3 closing epochs (650/550) were "fixed a priori, not tuned," but gives no analogous account for the 120°/150° gate or the 0.05 magnitude floor, and §7 acknowledges none of these is derived. Since the gate is the load-bearing mechanism distinguishing GradBlend from the failed peer rules, a threshold-sensitivity ablation (e.g., gate at 100°/130°, 110°/140°, 130°/160°, and magnitude floor 0.01/0.1) on the voxel loss under S3 is needed to show the result is not an artifact of thresholds placed on top of the observed conflict angle. Similarly, ρ is fixed to 1 with no ablation; a ρ ∈ {0.25, 0.5, 2} sweep, or a justification, would strengthen §4.2.
  4. [§5.1, Eq. (15); Appendix A.6] CFD is introduced as a metric and used as evidence, but its own properties are not characterized. Table 1 reports CFD for the prior surrogates as single values with no uncertainty; since CFD is a sample statistic over 100k showers, its sampling variability (e.g., bootstrap over showers) should be reported at least once to calibrate what differences are meaningful — this directly bears on whether the 0.002 Base-vs-Lantern gap and the 0.113 CaloDREAM-vs-0.120 Lantern gap are resolvable at all. Relatedly, there is no demonstration that CFD responds to correlation errors it is designed to catch (e.g., a synthetic corruption of the reference that preserves marginals but scrambles inter-layer correlations, showing FPD/KPD insensitive and CFD sensitive). Without this, the claim that CFD captures structure "FPD and KPD miss" (§1, contributions) is asserted rather than shown.
  5. [§5 (Experimental Setup); §7] The generalization claim "our method carries over to the finer Dataset 3 ... without modification" (§5) is stated with no supporting experiment and should be removed or explicitly marked as conjecture. More broadly, all evidence comes from a single dataset, geometry, and backbone (Dataset 2, ViT+DDPM). This is acceptable for a first controlled study — the within-paper comparisons are clean — but the abstract's framing ("admits the same signal without regression") reads as a general statement about the combination rule. The authors should either add one transfer experiment (the LEMURS geometries or Dataset 3 are named in §7 as "direct targets") or scope the claims to the demonstrated setting throughout the abstract and §6.
minor comments (8)
  1. [§5.3 vs Table 1] Text says Base reaches "a CFD of 0.1215" while Table 1 reports 0.122±0.004. Please make consistent (state whether 0.1215 is a rounded mean or a specific seed).
  2. [Table 1] Base+IMTL-G (Laplacian) row: FPD 3454±4820 and KPD 23.520±34.500 have std exceeding the mean, indicating at least one diverged seed (acknowledged in text). Please state explicitly how many of the three seeds diverged and whether means include the diverged run; likewise for PCGrad-Laplacian (169.0±72.8).
  3. [§4.1.2, Eq. (9)] λ_max(L) ≈ 9.77 for this grid is cited as "computed once and cached"; for reproducibility, state whether this is the exact spectral norm of the 6480-node combinatorial Laplacian and note the graph construction (6-connectivity, periodic φ) determines it independent of data.
  4. [§4.3] The cancellation of w_aux(s) from the GradBlend update is a nice observation but is stated only in prose; a one-line derivation (unit normalization removes the positive scalar from ĝ_a; magnitude anchored to ||g_d||) would make it rigorous. Also clarify that for the peer baselines w_aux(s) does not cancel, so the "same schedule" fairness claim holds at the loss level but the baselines see a different effective objective scaling — this is inherent to the methods but worth one sentence.
  5. [§2; Appendix C.1] The GCS baseline family (Du et al. [13], Yan et al. [50]) is discussed as the closest prior art but not run as a baseline; since GradBlend is positioned against scalar-weight gating, a GCS row in Table 1 would substantially strengthen the ablation. At minimum, state why it was excluded.
  6. [Figure 1] The y-axes are unweighted auxiliary loss, but the Base curve (which never optimizes the auxiliary) is the key reference; consider annotating that Base's curve is a pure evaluation pass. Legend colors for GradNorm vs ConFIG are hard to distinguish in grayscale.
  7. [Appendix D.2, Table 7] The scheduled-UW rows (FPD 115–149) are far above Base (30.9) yet the text calls sched UW "competitive" and "strongest" among loss-weighting methods; the qualifier "among loss-weighting baselines" should be kept wherever this is mentioned to avoid confusing readers skimming the appendix.
  8. [References] Several 2026-dated arXiv entries ([23], [32], [46], [47]) should be checked for venue/version updates; [41] (CaloDiT) is cited as an ACAT presentation — add the publication if one now exists. The GitHub link in footnote 2 mentions "kdd27" — verify this is the intended permanent repository.

Circularity Check

0 steps flagged

No significant circularity: empirical ML methods paper evaluated on external Geant4/CaloChallenge benchmarks; nothing reduces to its inputs by construction.

full rationale

Lantern proposes two auxiliary losses, a gradient-combination rule (GradBlend), and a correlation metric (CFD), then measures them against held-out Geant4 showers and public CaloChallenge submissions. CFD is a normalized Frobenius distance between Pearson correlation tensors of generated vs. reference samples—an evaluation score, not a quantity derived from a fitted input. The voxel residual and Laplacian losses compare the model’s clean estimate to detached Geant4 references; they do not redefine success as their own optimum. GradBlend is an engineered primary-anchored update (unit-direction blend, magnitude anchored to ∥g_d∥, angular gate); the paper does not claim it is forced by a uniqueness theorem or first-principles derivation. Preprocessing that fixes per-layer energies is stated explicitly and inherited from CaloDREAM, not hidden as a predicted conservation law. Hyperparameters (120°/150° gates, schedule close epochs) are empirical design choices evaluated by ablation, not fitted targets relabeled as predictions. Self-citations (e.g., authors’ prior evaluation survey) are peripheral, not load-bearing for the central claims. The skeptic’s concern that the FPD/CFD “improvement” over Base is thin is a statistical-power/correctness issue, not circularity. Derivation chain is self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 5 invented entities

Empirical ML methods paper: relies on standard diffusion/multi-task math, Geant4-as-truth domain assumption, and several hand-set optimization thresholds that the central stability claim depends on. No new physical entities; invented items are methodological constructs (CFD, GradBlend, losses).

free parameters (6)
  • GradBlend conflict gate thresholds (120° full, 150° cutoff) = 120° / 150°
    Direction gate breakpoints chosen empirically; load-bearing for when auxiliary steers vs is suppressed (§4.2, §7).
  • GradBlend magnitude floor = 0.05
    Prevents step collapse under near-antiparallel gradients; fixed, not derived (§4.2).
  • Auxiliary schedule closing epochs (S3) = epoch 650 (voxel), 550 (Laplacian)
    Terminal denoising-only phase epochs fixed a priori per loss type; essential for voxel loss sample quality (§5.4).
  • Voxel residual Huber δ and energy floor ε_vox = δ_vox=1, ε_vox=1e-6
    Robust loss hyperparameters for variance-stabilized residuals (§4.1.1).
  • Laplacian scale 1/λ_max(L)^2 = λ_max(L)≈9.77
    Operator-norm normalization from cached λ_max≈9.77 so auxiliary gradient scale matches denoising assumption (§4.1.2).
  • GradBlend ρ (tentative blend coefficient) = 1
    Fixed auxiliary share before gating; set to 1 for unit bisector (§4.2).
axioms (7)
  • domain assumption Denoising diffusion (DDPM) noise-prediction objective is a valid primary generative training signal for shower distributions.
    Entire primary-task framing and Base baseline rest on standard DDPM theory (§3, Eq. 3).
  • domain assumption Geant4 samples define the target physical distribution; matching them (FPD/KPD/CFD/classifier) is the success criterion.
    Standard CaloChallenge assumption; all metrics compare to held-out Geant4 (§5).
  • domain assumption No closed-form per-sample PDE or hard voxel-wise conservation law is available; only soft structure (counting fluctuations, local geometry differences) can supervise.
    Motivates auxiliaries vs classical PINN residuals (§1–2).
  • ad hoc to paper Denoising must remain strictly primary; peer treatment of generative and physics gradients is inappropriate under conflict.
    Design axiom of GradBlend; supported empirically here but not proved universal (§3, §4.2).
  • domain assumption Shower voxel fluctuations scale approximately as sqrt(energy) (counting statistics), justifying inverse-variance residual weights.
    Cited calorimetry background for ℓ_vox (§4.1.1, Fabjan & Gianotti).
  • domain assumption Face-adjacent grid graph Laplacian responses capture relevant local shower structure to match.
    Defines ℓ_lap; showers are not smoothed—mismatch of high-pass responses is penalized (§4.1.2).
  • ad hoc to paper Unit-vector direction blend plus denoising-norm magnitude yields a stable multi-objective step when angles are gated as specified.
    Core algorithmic hypothesis; Algorithm 1; earlier norm-coupled variants said unstable (§4.2).
invented entities (5)
  • Correlation Frobenius Distance (CFD) independent evidence
    purpose: Single normalized score for layer-wise and voxel-wise correlation fidelity missing from FPD/KPD/PCC matrices.
    Defined in §1 and Eqs. (14)–(17); computed on external Geant4 vs generated correlations—methodological metric, falsifiable by disagreement with other correlation probes.
  • GradBlend update rule no independent evidence
    purpose: Combine denoising and auxiliary gradients while anchoring magnitude to denoising and gating directional share by conflict angle.
    Algorithm 1; new combination operator relative to cited MTL methods.
  • Variance-stabilized voxel residual loss ℓ_vox no independent evidence
    purpose: Soft per-voxel physics auxiliary from counting-stat scaled Huber residuals on clean estimate vs reference.
    §4.1.1; adaptation of known robust/weighted residuals to showers.
  • Graph-Laplacian mismatch loss ℓ_lap no independent evidence
    purpose: Penalize difference of Laplacian responses between prediction and Geant4 on detector adjacency graph.
    §4.1.2; not a smoothness regularizer but a structure-matching term.
  • Lantern (Base+GradBlend physics-guided surrogate) no independent evidence
    purpose: Named system combining DDPM shape network, auxiliaries, GradBlend, and schedule S3.
    Productization of the above on CaloChallenge DS2.

pith-pipeline@v1.2.0-grok45-kimik3 · 29897 in / 4378 out tokens · 74736 ms · 2026-07-31T02:19:04.625350+00:00 · methodology

0 comments
read the original abstract

Monte Carlo simulation of calorimeter showers is a principal bottleneck for the High-Luminosity LHC, and diffusion models have emerged as fast, high-fidelity surrogates. Their denoising objective is purely statistical, however: a model can minimize it while placing the physics wrong. Existing physics-informed generative methods cannot close this gap, because they assume a closed-form law, a governing PDE residual or a hard per-sample constraint, that a shower does not supply: no per-sample PDE governs a stochastic cascade, and energy conservation fixes only one scalar per shower. Standard metrics ignore the correlation structure across calorimeter layers and voxels, comparing showers only in a physics feature space. We address both gaps. We introduce the Correlation Frobenius Distance (CFD), a single normalized score for correlation fidelity at layer-wise and voxel-wise scales. We then encode the soft per-sample structure available in a shower as two physics-aware auxiliary losses: a variance-stabilized voxel residual loss grounded in counting statistics, and a graph Laplacian loss over the detector geometry. We combine both with denoising through GradBlend, which anchors the step magnitude to the denoising gradient while letting the auxiliary steer its direction, yielding Lantern, a physics-guided diffusion surrogate. On CaloChallenge Dataset 2, injecting the physics losses through task-symmetric rules such as PCGrad, GradNorm, IMTL-G, and ConFIG inflates FPD by 2-100x relative to denoising alone, whereas GradBlend admits the same signal without regression and, with the Laplacian loss, Lantern improves both FPD and CFD. Our ablation on the auxiliary loss scheduler shows that the voxel residual loss, whose gradient conflicts with denoising, requires a terminal denoising-only phase to preserve shower fidelity, whereas the non-conflicting Laplacian loss is insensitive to the schedule.

Figures

Figures reproduced from arXiv: 2607.25060 by Farzana Yasmin Ahmad, Geoffrey Fox, Vanamala Venkataswamy.

Figure 1
Figure 1. Figure 1: Unweighted auxiliary loss over training on [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 2 canonical work pages

  1. [1]

    Agostinelli et al

    S. Agostinelli et al. 2003. GEANT4—A Simulation Toolkit.Nuclear Instruments and Methods in Physics Research A506, 3 (2003), 250–303

    2003

  2. [2]

    Farzana Yasmin Ahmad, Vanamala Venkataswamy, and Geoffrey Fox. 2024. A comprehensive evaluation of generative models in calorimeter shower simulation. arXiv preprint arXiv:2406.12898(2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Oz Amram and Kevin Pedro. 2023. Denoising Diffusion Models with Geometry Adaptation for High Fidelity Calorimeter Simulation.Physical Review D108 (2023), 072014

    2023

  4. [4]

    Giorgio Apollinari, O Brüning, Tatsushi Nakamoto, and Lucio Rossi. 2017. High luminosity large hadron collider HL-LHC.arXiv preprint arXiv:1705.08830(2017)

    Pith/arXiv arXiv 2017

  5. [5]

    Giacomo Baldan, Qiang Liu, Alberto Guardone, and Nils Thuerey. 2025. Flow matching meets pdes: A unified framework for physics-constrained generation. arXiv preprint arXiv:2506.08604(2025)

    arXiv 2025

  6. [6]

    Kochmann

    Jan-Hendrik Bastek, WaiChing Sun, and Dennis M. Kochmann. 2025. Physics- Informed Diffusion Models. InInternational Conference on Learning Representa- tions (ICLR). arXiv:2403.14404 [cs.LG]

    Pith/arXiv arXiv 2025

  7. [7]

    Joschka Birk, Anna Hallin, and Gregor Kasieczka. 2024. OmniJet-𝛼: The first cross-task foundation model for particle physics.Machine Learning: Science and Technology5, 3 (2024), 035031. doi:10.1088/2632-2153/ad66ad

    doi:10.1088/2632-2153/ad66ad 2024

  8. [8]

    Erik Buhmann, Frank Gaede, Gregor Kasieczka, Anatolii Korol, William Korcari, Katja Krüger, and Peter McKeown. 2024. CaloClouds II: Ultra-Fast Geometry- Independent Highly-Granular Calorimeter Simulation.Journal of Instrumentation 19 (2024), P04020

    2024

  9. [9]

    CaloChallenge Collaboration. 2022. Fast Calorimeter Simulation Challenge 2022 – Homepage and Evaluation Code. https://github.com/CaloChallenge/homepage. Accessed: 2026-07-13

    2022

  10. [10]

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multi- task Networks. InProceedings of the 35th International Conference on Machine Learning (ICML) (Proceedings of Machine Learning Research, Vol. 80). PMLR, 794– 803

    2018

  11. [11]

    McCann, Marc L

    Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, and Jong Chul Ye. 2023. Diffusion Posterior Sampling for General Noisy Inverse Problems. InInternational Conference on Learning Representations (ICLR)

    2023

  12. [12]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Pith/arXiv arXiv 2020

  13. [13]

    Czarnecki, Siddhant M

    Yunshu Du, Wojciech M. Czarnecki, Siddhant M. Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. 2018. Adapting Auxiliary Losses Using Gradient Similarity. arXiv:1812.02224

    Pith/arXiv arXiv 2018

  14. [14]

    Christian W Fabjan and Fabiola Gianotti. 2003. Calorimetry for particle physics. Reviews of Modern Physics75, 4 (2003), 1243

    2003

  15. [15]

    2022.Fast Calorimeter Simulation Challenge 2022 - Dataset 2

    Michele Faucci Giannelli, Gregor Kasieczka, Claudius Krause, Ben Nachman, Dalila Salamani, David Shih, and Anna Zaborowska. 2022.Fast Calorimeter Simulation Challenge 2022 - Dataset 2. doi:10.5281/zenodo.6366271

    work page doi:10.5281/zenodo.6366271 2022

  16. [16]

    2022.Fast Calorimeter Simulation Challenge 2022 - Dataset 3

    Michele Faucci Giannelli, Gregor Kasieczka, Claudius Krause, Ben Nachman, Dalila Salamani, David Shih, and Anna Zaborowska. 2022.Fast Calorimeter Simulation Challenge 2022 - Dataset 3. doi:10.5281/zenodo.6366324

    work page doi:10.5281/zenodo.6366324 2022

  17. [17]

    2025.Fast Calorimeter Simulation Challenge 2022 - Submissions Dataset 2

    Michele Faucci Giannelli, Gregor Kasieczka, Claudius Krause, Benjamin Nach- man, Dalila Salamani, David Shih, Anna Zaborowska, Oz Amram, Kerstin Bor- ras, Matthew Buckley, Thorsten Buss, Renato Paulo Da Costa Cardoso, Vijay Ekambaram, Florian Ernst, Luigi Favaro, Frank Gaede, Shih-Chieh Hsu, Kristina Jaruskova, Benno Käch, Jayant Kalagnanam, Dirk Krücker,...

    2025

  18. [18]

    Luigi Favaro, Ayodele Ore, Sofia Palacios Schweitzer, and Tilman Plehn. 2025. CaloDREAM: Detector Response Emulation via Attentive Flow Matching.SciPost Physics18 (2025), 088

    2025

  19. [19]

    Maddix, Yi Zhu, Mu Li, and Yuyang Wang

    Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle C. Maddix, Yi Zhu, Mu Li, and Yuyang Wang. 2023. PreDiff: Precipitation Nowcast- ing with Latent Diffusion Models. InAdvances in Neural Information Processing Systems (NeurIPS)

    2023

  20. [20]

    Majdi Hassan et al. 2024. ET-Flow: Equivariant Flow-Matching for Molecular Conformer Generation. InAdvances in Neural Information Processing Systems (NeurIPS). arXiv:2410.22388

    Pith/arXiv arXiv 2024

  21. [21]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. arXiv:2006.11239 [cs.LG]

    Pith/arXiv arXiv 2020

  22. [22]

    Emiel Hoogeboom, Victor Garcia Satorras, Clément Vignac, and Max Welling

  23. [23]

    Cheng Jiang, Sitian Qian, Kevin Pedro, Oz Amram, Huilin Qu, and Maggie Voetberg. 2026. CaloTrilogy: Toward a Breakthrough in One-Step, End-to-End, Physics-Guided Shower Generation for Modern Calorimeters.arXiv preprint arXiv:2606.04165(2026)

    Pith/arXiv arXiv 2026

  24. [24]

    Raghav Kansal, Anni Li, Javier Duarte, Nadezda Chernyavskaya, Maurizio Pierini, Breno Orzari, and Thiago Tomei. 2023. Evaluating generative models in high energy physics.Physical Review D107, 7 (2023), 076017

    2023

  25. [25]

    Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. InProceedings of the IEEE conference on computer vision and pattern recognition. 7482–7491

    2018

  26. [26]

    Dmitrii Kobylianskii, Nathalie Soybelman, Etienne Dreyer, and Eilam Gross

  27. [27]

    Claudius Krause, Michele Faucci Giannelli, Gregor Kasieczka, Benjamin Nach- man, Dalila Salamani, David Shih, Anna Zaborowska, Oz Amram, Kerstin Borras, Matthew R Buckley, et al. 2025. CaloChallenge 2022: a community challenge for fast calorimeter simulation.Reports on Progress in Physics88, 11 (2025), 116201

    2025

  28. [28]

    Claudius Krause and David Shih. 2021. CaloFlow II: Even faster and still accu- rate generation of calorimeter showers with normalizing flows.arXiv preprint arXiv:2110.11377(2021)

    Pith/arXiv arXiv 2021

  29. [29]

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. 2019. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems32 (2019)

    2019

  30. [30]

    Matthew Leigh, Debajyoti Sengupta, Guillaume Quétant, John Andrew Raine, Knut Zoch, and Tobias Golling. 2024. PC-JeDi: Diffusion for Particle Cloud Generation in High Energy Physics.SciPost Physics16 (2024), 018

    2024

  31. [31]

    Matthew Leigh, Debajyoti Sengupta, John Andrew Raine, Guillaume Quétant, and Tobias Golling. 2024. PC-Droid: Faster Diffusion and Improved Quality for Particle Cloud Generation.Physical Review D(2024). arXiv:2307.06836

    Pith/arXiv arXiv 2024

  32. [32]

    Yunyang Li, Lin Huang, Luojia Xia, et al . 2026. Elign: Equivariant Diffusion Model Alignment from Foundational Machine Learning Force Fields.arXiv preprint arXiv:2601.21985(2026). arXiv:2601.21985 [cs.LG]

    arXiv 2026

  33. [33]

    Jinhao Liang, Yixuan Sun, Anirban Samaddar, Sandeep Madireddy, and Ferdi- nando Fioretto. 2025. Chance-Constrained Flow Matching. InAdvances in Neural Information Processing Systems (NeurIPS). arXiv:2509.25157

    arXiv 2025

  34. [34]

    Liyang Liu, Yi Li, Zhanghui Kuang, Jing-Hao Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. 2021. Towards impartial multi-task learning. InInternational conference on learning representations

    2021

  35. [35]

    Qiang Liu, Mengyu Chu, and Nils Thuerey. 2025. ConFIG: Towards Conflict-free Training of Physics Informed Neural Networks. InInternational Conference on Learning Representations (ICLR). arXiv:2408.11104

    Pith/arXiv arXiv 2025

  36. [36]

    Vinicius Mikuni and Benjamin Nachman. 2022. Score-based generative models for calorimeter shower simulation.Physical Review D106, 9 (2022), 092009

    2022

  37. [37]

    Vinicius Mikuni and Benjamin Nachman. 2024. CaloScore v2: single-shot calorimeter shower simulation with diffusion models.Journal of Instrumen- tation19, 02 (2024), P02001. doi:10.1088/1748-0221/19/02/P02001

    doi:10.1088/1748-0221/19/02/p02001 2024

  38. [38]

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. 2020. Reliable fidelity and diversity metrics for generative models. InInternational conference on machine learning. PMLR, 7176–7185

    2020

  39. [39]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

    2023

  40. [40]

    Platt and Alan H

    John C. Platt and Alan H. Barr. 1988. Constrained Differential Optimization. In Neural Information Processing Systems (NIPS), Dana Z. Anderson (Ed.), Vol. 1. American Institute of Physics, 612–621

    1988

  41. [41]

    Piyush Raikwar, Renato Cardoso, Kristina Jaruskova, Dalila Salamani, Sofia Vallecorsa, Anna Zaborowska, et al. 2024. CaloDiT: Diffusion with transformers for fast shower simulation. Presentation at the 22nd International Workshop on Advanced Computing and Analysis Techniques in Physics Research (ACAT). https://indico.cern.ch/event/1330797/contributions/5796591/

    arXiv 2024

  42. [42]

    Piyush Raikwar, Anna Zaborowska, Peter McKeown, Renato Cardoso, Mikolaj Piorczynski, and Kyongmin Yeo. 2025. A Generalisable Generative Model for Multi-Detector Calorimeter Simulation.arXiv preprint arXiv:2509.07700(2025)

    Pith/arXiv arXiv 2025

  43. [43]

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The emerging field of signal processing on graphs: Extend- ing high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine30, 3 (2013), 83–98

    2013

  44. [44]

    Kingma, Abhishek Kumar, Ste- fano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Ste- fano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochastic Differential Equations. arXiv:2011.13456 [cs.LG]

    Pith/arXiv arXiv 2021

  45. [45]

    Qingwen Tian, Yuxin Xu, Yixuan Yang, Zhen Wang, Ziqi Liu, Pengju Yan, and Xiaolin Li. 2025. EquiFlow: Equivariant Conditional Flow Matching with Optimal Transport for 3D Molecular Conformation Prediction. InAAAI Conference on Artificial Intelligence

    2025

  46. [46]

    Noah Trupin, Rahul Ghosh, and Aadi Jangid. 2026. Learning Flow Distri- butions via Projection-Constrained Diffusion on Manifolds.arXiv preprint arXiv:2602.17773(2026). Ahmad et al

    arXiv 2026

  47. [47]

    Utkarsh Utkarsh, Pengfei Cai, Alan Edelman, Rafael Gomez-Bombarelli, and Christopher Rackauckas. 2026. Physics-constrained flow matching: Sampling generative models with hard constraints.Advances in Neural Information Pro- cessing Systems38 (2026), 160217–160252

    2026

  48. [48]

    Lorenz Vaitl and Leon Klein. 2025. Path Gradients after Flow Match- ing. InAdvances in Neural Information Processing Systems (NeurIPS). arXiv:2505.10139 [cs.LG]

    arXiv 2025

  49. [49]

    Sifan Wang, Yujun Teng, and Paris Perdikaris. 2021. Understanding and Miti- gating Gradient Flow Pathologies in Physics-Informed Neural Networks.SIAM Journal on Scientific Computing43, 5 (2021), A3055–A3081

    2021

  50. [50]

    Junjun Yan, Xinhai Chen, Zhichao Wang, Enqiang Zhou, and Jie Liu. 2023. Auxiliary-tasks learning for physics-informed neural network-based partial differential equations solving.arXiv preprint arXiv:2307.06167(2023)

    Pith/arXiv arXiv 2023

  51. [51]

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient Surgery for Multi-Task Learning. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 33. 5824–5836

    2020

  52. [52]

    Yi Zhang and Difan Zou. 2025. Physics-informed distillation of diffusion models for pde-constrained generation.arXiv preprint arXiv:2505.22391(2025). A Evaluation Protocol and Metric Details A.1 Samples and Reference Data All metrics are computed against the second file of CaloChallenge Dataset 2, which contains 100,000 Geant4 electron showers held out fr...

    Pith/arXiv arXiv 2025

  53. [2022]

    InInternational Conference on Machine Learning (ICML)

    Equivariant Diffusion for Molecule Generation in 3D. InInternational Conference on Machine Learning (ICML). Oral presentation

  54. [2024]

    arXiv:2402.11575

    CaloGraph: Graph-Based Diffusion Model for Fast Shower Genera- tion in Calorimeters with Irregular Geometry.Physical Review D(2024). arXiv:2402.11575

    Pith/arXiv arXiv 2024