pith. machine review for the scientific record. sign in

arxiv: 2604.19846 · v1 · submitted 2026-04-21 · ✦ hep-ex · astro-ph.HE· astro-ph.IM· cs.AI· cs.LG

Recognition: unknown

Neural posterior estimation of the neutrino direction in IceCube using transformer-encoded normalizing flows on the sphere

A. A. Harnisch, A. Balagopal V., A. Chubarov, A. Connolly, A. Domi, A. Eimer, A. Fedynitch, A. Franckowiak, A. Garcia, A. Ghadimi, A. Granados, A. Hallgren, A. Haungs, A. Hidvegi, A. Hollnagel, A. Ishihara, A. Kappes, A. Karle, A. Katil, A. Khatee Zathul, A. Kheirandish, A. Kochocki, A. Kravka, A. Kumar, A. Leszczy\'nska, A. Mand, A. Medina, A. Mosbrugger, A. Noell, A. Novikov, A. Obertacke, A. Olivas, A. Parenti, A. Pont\'en, A. Rehman, A. R. Fazely, A. Rifaie, A. Rosted, A. Sandrock, A. Scholz, A. Terliuk, A. Thakuri, A. Vaidyanathan, A. Vijai, A. Wang, A. Weindl, A. Y. Wen, A. Zander Jurowitzki, A. Zegarelli, B. A. Clark, B. Brinson, B. Henke, B. Owens, B. Pries, B. Riedel, B. Schl\"uter, B. Skrzypek, C. Arg\"uelles, C. Bellenghi, C. Boscolo Meneguolo, C. De Clercq, C. D. Rho, C. Eldridge, C. Finley, C. Glaser, C. G\"unther, C. Ha, C. Hill, C. H. Wiebusch, C. Klein, C. Kopper, C. Lagunas Gualda, C. Li, C. Lin, C. Love, C. P\'erez de los Heros, C. Raab, C. Rott, C. Spannfellner, C. Spiering, C. Walck, C. Weaver, C. Wendt, D. A. Coloma Borja, D. Berley, D. Butterfield, D. Chirkin, D. Delgado, D. Durnford, D. Els\"asser, D. F. Cowen, D. Fox, D. Grant, D. Guevel, D. Hooper, D. J. Koskinen, D. Kang, D. Mousadi, D. R. Williams, D. Ryckbosch, D. Salazar-Gallegos, D. Seckel, D. Soldin, D. Tosi, D. Veske, D. Z. Besson, E. Bernardini, E. Blaufuss, E. Ellinger, E. Ganster, E. Genton, E. H. S. Warrick, E. J. Roberts, E. Krupczak, E. Kun, E. Magnus, E. Manao, E. Moyaux, E. N. Paudel, E. O'Sullivan, E. Resconi, E. Yildizci, F. Bontempo, F. G. Schr\"oder, F. Halzen, F. Henningsen, F. Kirchner, F. Lauber, F. Lucarelli, F. Mayhew, F. McNally, F. Schl\"uter, F. Varsi, F. Yu, G. C. Hill, G. de Wasseige, G. H. Collin, G. M. Spiczak, G. Sanger-Johnson, G. Sommani, G. T. Przybylski, G. Wrede, G. W. Sullivan, H. Erpenbeck, H. Kimku, H. Kolanoski, H. Niederhausen, H. Pandya, H. Schieler, I. C. Mari\c{s}, I. Martinez-Soler, I. Reistroffer, I. Taboada, J. A. Aguilar, J. Adams, J. A. Torres, J. Becker Tjus, J. Beise, J. B\"ottcher, J. Braun, J. Carpio, J. C. D\'iaz-V\'elez, J. Evans, J. Gallagher, J. G. Gonzalez, J. Hardin, J. H\"au{\ss}ler, J. Hellrung, J. J. Beatty, J. J. DeLaunay, J. Kiryluk, J. Liao, J. L. Kelley, J.M. Alameddine, J. Mauro, J. M. Conrad, J. Mitchell, J. Necker, J. Osborn, J. Peterson, J. P. Lazar, J. P. Yanez, J. Rack-Helleis, J. Saffer, J. Soedingrekso, J. Stachurska, J. Thwaites, J. Valverde, J. Vandenbroucke, J. van Santen, J. Vara, J. Villarreal, J. Weldert, J. Werthebach, J. Y. Book Motzkin, K. Andeen, K. Carloni, K. D. de Vries, K. D. Hoffman, K. Dutta, K. Fang, K. Farrag, K. Hanson, K. Helbing, K. Hoshina, K. Hultqvist, K. Hymon, K. Kruiswijk, K. Leonard DeHolton, K. L. Fan, K. Meagher, K. M. Groth, K. Noda, K. Rawlins, K. Tollefson, K. Upshaw, L. Bloom, L. Draper, L. Dueser, L. Eidenschink, L. Gerhardt, L. Halve, L. Hamacher, L. Hennig, L. Heuermann, L. Kardum, L. K\"opke, L. Lallement Arnaud, L. Lu, L. Marten, L. Merten, L. Molchany, L. Neste, L. Paul, L. Pyras, L. Ricca, L. Ruohan, L. Schlickmann, L. Seen, L. Van Rootselaar, L. Witthaus, M. Ackermann, M. A. DuVernois, M. Ahlers, M. Dittmer, M. Garcia, M. Handt, M. Hostert, M. Hrywniak, M. Jacquart, M. Jansson, M. Jin, M. J. Larson, M. Kauer, M. Khanal, M. Kowalski, M. Liubarska, M. Macdonald, M. Meier, M. Nakos, M. Neumann, M. Plum, M. Ravn, M. Rongen, M. Santander, M. Scarnera, M. Schaufel, M. Seikh, M. Stamatikos, M. Thiesmeyer, M. U. Nisa, M. Venugopal, M. Vereecken, M. Weyrauch, N. Chau, N. Feigl, N. Heyer, N. Kamp, N. Krieger, N. Kurahashi, N. M. Amin, N. Park, N. Rad, N. Schmeisser, N. Shimizu, N. Valtonen-Mattila, N. van Eijndhoven, N. Whitehorn, O. Botner, O. Janik, P. A. Evenson, P. A. Sevle Myhr, P. Behrens, P. Desiati, P. Eller, P. F\"urst, P. Gutjahr, P. Hatch, P. Koundal, P. Sampathkumar, P. Soldin, P. Weigel, P. Zhelnin, P. Zilberman, Q. R. Liu, R. Abbasi, R. Babu, R. Bay, R. Engel, R. Hewett, R. Hmaid, R. Koirala, R. Maruyama, R. Orsoe, R. Procter-Murphy, R. Shah, R. Snihur, R. T. Burley, R. W. Moore, R. Young, S. Ali, S. Athanasiadou, S. Benkel, S. BenZvi, S. Blot, S. B\"oser, S. Choi, S. Deng, S. DiKerby, S. Eulig, S. Fukami, S. Goswami, S. Griffin, S. Hickford, S. Hori, S. Jain, S. J. Gray, S. Koch, S. Mancina, S. Marka, S. Mondal, S. N. Axani, S. Pick, S. Reusch, S. R. Klein, S. Rodan, S. Sarkar, S. Schindler, S. Schwirn, S. Sclafani, S. Seunarine, S. Shah, S. Shefali, S. Ter-Antonyan, S. Tilav, S. Toscano, S. Vergara Carrasco, S. Verpoest, S. W. Barwick, S. Yoshida, S. Yu, S. Yun-C\'arcamo, S. Zhang, T. C. Petersen, T. Delmeulle, T. DeYoung, T. Ding, T. Ehrhardt, T. Gl\"usenkamp, T. Huber, T. Karg, T. Kim, T. Kontrimas, T. Kozynets, T. Krishnan, T. Montaruli, T. Mukherjee, T. Pernice, T. Ruhe, T. Schmidt, T. Stanev, T. Stezelberger, T. St\"urwald, T. Stuttard, T. Van Eeden, T. Yuan, U. Naumann, V. Basu, V. O'Dell, V. Palusova, V. Parrish, V. Poojyam, W. Esmail, W. G. Thompson, W. Hou, W. Iwakiri, W. Kang, W. Luszczak, W. Rhode, X. Bai, X. W. Xu, Y. Ashida, Y. Kobayashi, Y. Lyu, Y. Makino, Y. Merckx, Y. Morii, Y. T. Liu, Y. Yao, Z. Brisson-Tsavoussis, Z. Chen, Z. Marka, Z. Rechav, Z. Zhang

Pith reviewed 2026-05-10 01:11 UTC · model grok-4.3

classification ✦ hep-ex astro-ph.HEastro-ph.IMcs.AIcs.LG
keywords neutrino direction reconstructionIceCubeneural posterior estimationnormalizing flowstransformer encoderangular resolutiontracksshowers
0
0 comments X

The pith

A transformer encoder feeding a spherical normalizing flow reconstructs neutrino directions in IceCube more accurately and faster than B-spline likelihood methods.

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

The paper develops a neural posterior estimator that uses a transformer to map IceCube event hits directly to the parameters of a normalizing flow defined on the two-sphere. This produces full posterior distributions over neutrino arrival directions for both track-like and shower-like events. At 100 TeV the median angular resolution improves by factors of 1.3 for throughgoing tracks, 1.7 for showers, and 2.5 for starting tracks relative to current likelihood fits. The computation time remains constant regardless of how broad or narrow the posterior is, enabling all-sky scans in seconds rather than hours. A reader would care because sharper and quicker direction estimates increase the chance of linking high-energy neutrinos to their astrophysical sources.

Core claim

The central claim is that a transformer encoder with dual residual streams, nonlinear QKV projections, and a dedicated class token can predict the parameters of a novel spherical normalizing flow built from C²-smooth rational-quadratic splines, scale transformations, and rotations; when trained on simulated IceCube events this yields state-of-the-art median angular resolution across 100 GeV to 100 PeV for both tracks and showers, outperforming B-spline likelihood reconstructions by the factors noted above while maintaining constant and much lower run time.

What carries the argument

A transformer encoder whose output directly parametrizes a normalizing flow on the 2-sphere constructed from C²-smooth rational-quadratic splines together with scale and rotation transformations.

If this is right

  • All-sky neutrino source searches become feasible on timescales of seconds instead of hours.
  • Real-time directional reconstruction is practical for high-energy neutrino alerts.
  • The method maintains its speed advantage even when the posterior spans the entire sky.
  • Improved angular precision directly raises the significance of associations between neutrinos and candidate sources.
  • The same architecture works uniformly for both narrow and broad posteriors without retuning.

Where Pith is reading between the lines

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

  • The constant-time property could allow the reconstruction to be embedded directly in the detector's online trigger pipeline.
  • Similar spherical-flow heads might be attached to other transformer backbones for directional inference in radio or optical astronomy.
  • Joint modeling of direction and energy could be added by expanding the flow to a higher-dimensional manifold without changing the encoder.
  • If the simulation-reality gap proves small, the approach could reduce the need for repeated expensive likelihood evaluations in future analyses.

Load-bearing premise

The simulated events used for training and testing faithfully reproduce the real detector response, ice optical properties, and event morphologies with no large domain shift when the model is applied to actual observations.

What would settle it

Run the trained model on a set of real IceCube events whose true directions are independently known or can be cross-checked through multi-messenger coincidences, then compare the reported angular errors against those obtained from the standard B-spline likelihood fit on the same events.

Figures

Figures reproduced from arXiv: 2604.19846 by A. A. Harnisch, A. Balagopal V., A. Chubarov, A. Connolly, A. Domi, A. Eimer, A. Fedynitch, A. Franckowiak, A. Garcia, A. Ghadimi, A. Granados, A. Hallgren, A. Haungs, A. Hidvegi, A. Hollnagel, A. Ishihara, A. Kappes, A. Karle, A. Katil, A. Khatee Zathul, A. Kheirandish, A. Kochocki, A. Kravka, A. Kumar, A. Leszczy\'nska, A. Mand, A. Medina, A. Mosbrugger, A. Noell, A. Novikov, A. Obertacke, A. Olivas, A. Parenti, A. Pont\'en, A. Rehman, A. R. Fazely, A. Rifaie, A. Rosted, A. Sandrock, A. Scholz, A. Terliuk, A. Thakuri, A. Vaidyanathan, A. Vijai, A. Wang, A. Weindl, A. Y. Wen, A. Zander Jurowitzki, A. Zegarelli, B. A. Clark, B. Brinson, B. Henke, B. Owens, B. Pries, B. Riedel, B. Schl\"uter, B. Skrzypek, C. Arg\"uelles, C. Bellenghi, C. Boscolo Meneguolo, C. De Clercq, C. D. Rho, C. Eldridge, C. Finley, C. Glaser, C. G\"unther, C. Ha, C. Hill, C. H. Wiebusch, C. Klein, C. Kopper, C. Lagunas Gualda, C. Li, C. Lin, C. Love, C. P\'erez de los Heros, C. Raab, C. Rott, C. Spannfellner, C. Spiering, C. Walck, C. Weaver, C. Wendt, D. A. Coloma Borja, D. Berley, D. Butterfield, D. Chirkin, D. Delgado, D. Durnford, D. Els\"asser, D. F. Cowen, D. Fox, D. Grant, D. Guevel, D. Hooper, D. J. Koskinen, D. Kang, D. Mousadi, D. R. Williams, D. Ryckbosch, D. Salazar-Gallegos, D. Seckel, D. Soldin, D. Tosi, D. Veske, D. Z. Besson, E. Bernardini, E. Blaufuss, E. Ellinger, E. Ganster, E. Genton, E. H. S. Warrick, E. J. Roberts, E. Krupczak, E. Kun, E. Magnus, E. Manao, E. Moyaux, E. N. Paudel, E. O'Sullivan, E. Resconi, E. Yildizci, F. Bontempo, F. G. Schr\"oder, F. Halzen, F. Henningsen, F. Kirchner, F. Lauber, F. Lucarelli, F. Mayhew, F. McNally, F. Schl\"uter, F. Varsi, F. Yu, G. C. Hill, G. de Wasseige, G. H. Collin, G. M. Spiczak, G. Sanger-Johnson, G. Sommani, G. T. Przybylski, G. Wrede, G. W. Sullivan, H. Erpenbeck, H. Kimku, H. Kolanoski, H. Niederhausen, H. Pandya, H. Schieler, I. C. Mari\c{s}, I. Martinez-Soler, I. Reistroffer, I. Taboada, J. A. Aguilar, J. Adams, J. A. Torres, J. Becker Tjus, J. Beise, J. B\"ottcher, J. Braun, J. Carpio, J. C. D\'iaz-V\'elez, J. Evans, J. Gallagher, J. G. Gonzalez, J. Hardin, J. H\"au{\ss}ler, J. Hellrung, J. J. Beatty, J. J. DeLaunay, J. Kiryluk, J. Liao, J. L. Kelley, J.M. Alameddine, J. Mauro, J. M. Conrad, J. Mitchell, J. Necker, J. Osborn, J. Peterson, J. P. Lazar, J. P. Yanez, J. Rack-Helleis, J. Saffer, J. Soedingrekso, J. Stachurska, J. Thwaites, J. Valverde, J. Vandenbroucke, J. van Santen, J. Vara, J. Villarreal, J. Weldert, J. Werthebach, J. Y. Book Motzkin, K. Andeen, K. Carloni, K. D. de Vries, K. D. Hoffman, K. Dutta, K. Fang, K. Farrag, K. Hanson, K. Helbing, K. Hoshina, K. Hultqvist, K. Hymon, K. Kruiswijk, K. Leonard DeHolton, K. L. Fan, K. Meagher, K. M. Groth, K. Noda, K. Rawlins, K. Tollefson, K. Upshaw, L. Bloom, L. Draper, L. Dueser, L. Eidenschink, L. Gerhardt, L. Halve, L. Hamacher, L. Hennig, L. Heuermann, L. Kardum, L. K\"opke, L. Lallement Arnaud, L. Lu, L. Marten, L. Merten, L. Molchany, L. Neste, L. Paul, L. Pyras, L. Ricca, L. Ruohan, L. Schlickmann, L. Seen, L. Van Rootselaar, L. Witthaus, M. Ackermann, M. A. DuVernois, M. Ahlers, M. Dittmer, M. Garcia, M. Handt, M. Hostert, M. Hrywniak, M. Jacquart, M. Jansson, M. Jin, M. J. Larson, M. Kauer, M. Khanal, M. Kowalski, M. Liubarska, M. Macdonald, M. Meier, M. Nakos, M. Neumann, M. Plum, M. Ravn, M. Rongen, M. Santander, M. Scarnera, M. Schaufel, M. Seikh, M. Stamatikos, M. Thiesmeyer, M. U. Nisa, M. Venugopal, M. Vereecken, M. Weyrauch, N. Chau, N. Feigl, N. Heyer, N. Kamp, N. Krieger, N. Kurahashi, N. M. Amin, N. Park, N. Rad, N. Schmeisser, N. Shimizu, N. Valtonen-Mattila, N. van Eijndhoven, N. Whitehorn, O. Botner, O. Janik, P. A. Evenson, P. A. Sevle Myhr, P. Behrens, P. Desiati, P. Eller, P. F\"urst, P. Gutjahr, P. Hatch, P. Koundal, P. Sampathkumar, P. Soldin, P. Weigel, P. Zhelnin, P. Zilberman, Q. R. Liu, R. Abbasi, R. Babu, R. Bay, R. Engel, R. Hewett, R. Hmaid, R. Koirala, R. Maruyama, R. Orsoe, R. Procter-Murphy, R. Shah, R. Snihur, R. T. Burley, R. W. Moore, R. Young, S. Ali, S. Athanasiadou, S. Benkel, S. BenZvi, S. Blot, S. B\"oser, S. Choi, S. Deng, S. DiKerby, S. Eulig, S. Fukami, S. Goswami, S. Griffin, S. Hickford, S. Hori, S. Jain, S. J. Gray, S. Koch, S. Mancina, S. Marka, S. Mondal, S. N. Axani, S. Pick, S. Reusch, S. R. Klein, S. Rodan, S. Sarkar, S. Schindler, S. Schwirn, S. Sclafani, S. Seunarine, S. Shah, S. Shefali, S. Ter-Antonyan, S. Tilav, S. Toscano, S. Vergara Carrasco, S. Verpoest, S. W. Barwick, S. Yoshida, S. Yu, S. Yun-C\'arcamo, S. Zhang, T. C. Petersen, T. Delmeulle, T. DeYoung, T. Ding, T. Ehrhardt, T. Gl\"usenkamp, T. Huber, T. Karg, T. Kim, T. Kontrimas, T. Kozynets, T. Krishnan, T. Montaruli, T. Mukherjee, T. Pernice, T. Ruhe, T. Schmidt, T. Stanev, T. Stezelberger, T. St\"urwald, T. Stuttard, T. Van Eeden, T. Yuan, U. Naumann, V. Basu, V. O'Dell, V. Palusova, V. Parrish, V. Poojyam, W. Esmail, W. G. Thompson, W. Hou, W. Iwakiri, W. Kang, W. Luszczak, W. Rhode, X. Bai, X. W. Xu, Y. Ashida, Y. Kobayashi, Y. Lyu, Y. Makino, Y. Merckx, Y. Morii, Y. T. Liu, Y. Yao, Z. Brisson-Tsavoussis, Z. Chen, Z. Marka, Z. Rechav, Z. Zhang.

Figure 1
Figure 1. Figure 1: Schematic overview of transformer-based amortized neural posterior estimation for neutrino [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relation of amortized neural posterior estimation with normalizing flows (used here) to [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the constant-time skymap creation using normalizing flows irrespective of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The data encoding pipeline from photon hits to posterior prediction. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Inductive biases of a transformer encoding for a) the data encoding in posterior estimation [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Validation-loss curves for the best track and shower model. The x-axis shows optimization [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Test results for different hyperparameters for showers and throughgoing tracks. Models [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: On the left, angular resolution (16%, 50%, 84% quantiles) of the transformer normalizing flow (TNF) with and without saturated DOMs (TNF no sat.) compared to the state-of-the-art respective likelihood method based on B-splines, SplineMPEMax [9] for tracks and Taupede2024 [8] for showers. The improvement of median angular resolution of the new method versus the respective likelihood method is shown in the l… view at source ↗
Figure 9
Figure 9. Figure 9: Coverage split up in different deposited energy bands. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Data / Monte Carlo comparison of the zenith and azimuth of the maximum of the [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Data / Monte Carlo comparison of the zenith and azimuth of the maximum of the posterior [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Runtimes of TNF for CPU and GPU in comparison to the B-spline method which runs [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Contours (left) and event view (right) for example shower events. The contours are [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Contours (left) and event view (right) for example track events. The contours are indicated [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Depictions of rational quadratic splines with various constraints. Knots are shown as [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Conditioning function a = fa∗ (z) on the cylinder height z to produce a regularized parameter a to be used for the azimuthal angle flow function, rqsA(ϕ)a, depicted in Fig. 15d. At the poles (z = −1 and z = 1) the output is always a = fa∗ (z = −1/z = 1) = 0 which enforces an identity mapping as rqsA(ϕ|z)a=0 = ϕ. The horizontal dashed lines correspond to specific values of a ∗ . with p = − h0h1(w0 + w1) 2(… view at source ↗
Figure 17
Figure 17. Figure 17: Test results for different hyperparameters for starting tracks. Models are sorted by average [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
read the original abstract

IceCube is a cubic-kilometer-scale neutrino detector located at the geographic South Pole. A precise directional reconstruction of IceCube neutrinos is vital for associations with astronomical objects. In this context, we discuss neural posterior estimation of the neutrino direction via a transformer encoder that maps to a normalizing flow on the 2-sphere. It achieves a new state-of-the-art angular resolution for the two main event morphologies in IceCube - tracks and showers - while being significantly faster than traditional B-spline-based likelihood reconstructions. All-sky scans can be performed within seconds rather than hours, and take constant computation time, regardless of whether the posterior extent is arc-minutes or spans the whole sky. We utilize a combination of $C^2$-smooth rational-quadratic splines, scale transformations and rotations to define a novel spherical normalizing-flow distribution whose parameters are predicted as a whole as the output of the transformer encoder. We test several structural choices diverting from the vanilla transformer architecture. In particular, we find dual residual streams, nonlinear QKV projection and a separate class token with its own cross-attention processing to boost test-time performance. The angular resolution for both showers and tracks improves substantially over the whole trained energy range from 100 GeV to 100 PeV. At 100 TeV deposited energy, for example, the median angular resolution improves by a factor of $1.3$ for throughgoing tracks, by a factor of $1.7$ for showers and by a factor of $2.5$ for starting tracks compared to state-of-the art likelihood reconstructions based on B-splines. While previous machine-learning (ML) efforts have managed to obtain competitive shower resolutions, this is the first time an ML-based method outperforms likelihood-based muon reconstructions above 100 GeV.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a neural posterior estimation method for reconstructing neutrino directions in IceCube. A transformer encoder maps event data to the parameters of a novel spherical normalizing flow (using C²-smooth rational-quadratic splines, scale transformations, and rotations). The approach is reported to deliver median angular resolution improvements over B-spline likelihood reconstructions of 1.3× for throughgoing tracks, 1.7× for showers, and 2.5× for starting tracks at 100 TeV deposited energy, while enabling constant-time all-sky scans. Architectural variants (dual residual streams, nonlinear QKV projections, separate class token) are tested and shown to improve performance across 100 GeV–100 PeV.

Significance. If the performance gains hold under realistic conditions, the work would mark the first demonstration of an ML method outperforming likelihood-based muon reconstructions above 100 GeV in IceCube. The computational speed-up and the technical contribution of the spherical normalizing-flow construction could enable new analyses that were previously limited by reconstruction time or posterior complexity.

major comments (2)
  1. [Abstract] Abstract: The quoted resolution improvement factors (1.3×, 1.7×, 2.5× at 100 TeV) and the claim of being the first ML method to beat likelihood-based muon reconstructions above 100 GeV are obtained exclusively from Monte Carlo simulations generated with the same photon-propagation and ice model used to define the B-spline likelihood. No real-data validation, no comparison of posterior calibration on observed events, and no systematic variation of ice optical parameters are reported; this directly affects the load-bearing claim that the method outperforms state-of-the-art reconstructions in practice.
  2. [Results] Results and validation sections: The manuscript evaluates the transformer NPE on held-out simulated events but provides no quantitative assessment of domain-shift robustness (e.g., by retraining or testing under varied scattering/absorption lengths or hole-ice models). Because the network can exploit any mismatch between the shared simulation model and true detector response while the physics-based baseline cannot, the reported gains may be inflated without such tests.
minor comments (2)
  1. [Abstract] The abstract states that several transformer variants were tested but does not tabulate the exact hyperparameter settings or the quantitative ablation results for dual residual streams, nonlinear QKV, and class-token cross-attention; a supplementary table would improve reproducibility.
  2. [Methods] Notation for the spherical flow parameters (spline knots, scale factors, rotation matrices) is introduced without an explicit equation reference or diagram showing their composition; this reduces clarity for readers unfamiliar with spherical normalizing flows.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their careful and constructive review of our manuscript. We address each major comment below, acknowledging the simulation-based scope of the evaluation and making targeted revisions to clarify limitations and strengthen the presentation of results.

read point-by-point responses
  1. Referee: [Abstract] Abstract: The quoted resolution improvement factors (1.3×, 1.7×, 2.5× at 100 TeV) and the claim of being the first ML method to beat likelihood-based muon reconstructions above 100 GeV are obtained exclusively from Monte Carlo simulations generated with the same photon-propagation and ice model used to define the B-spline likelihood. No real-data validation, no comparison of posterior calibration on observed events, and no systematic variation of ice optical parameters are reported; this directly affects the load-bearing claim that the method outperforms state-of-the-art reconstructions in practice.

    Authors: We agree that the reported resolution improvements and the comparison to prior ML methods are derived exclusively from Monte Carlo simulations using the nominal ice model shared with the B-spline likelihood. This is the standard evaluation protocol in IceCube reconstruction papers, as true neutrino directions are unavailable for real events, precluding direct resolution measurements on data. The B-spline method is likewise defined and benchmarked within the same simulation framework, ensuring an internally consistent comparison. We have revised the abstract to explicitly qualify that the improvement factors apply to simulated events and to rephrase the 'first ML method' claim as 'the first ML-based method to outperform likelihood-based muon reconstructions on simulated events above 100 GeV.' We have also added coverage tests demonstrating posterior calibration on held-out simulations. Systematic variations of ice optical parameters and direct real-data tests lie outside the present scope. revision: partial

  2. Referee: [Results] Results and validation sections: The manuscript evaluates the transformer NPE on held-out simulated events but provides no quantitative assessment of domain-shift robustness (e.g., by retraining or testing under varied scattering/absorption lengths or hole-ice models). Because the network can exploit any mismatch between the shared simulation model and true detector response while the physics-based baseline cannot, the reported gains may be inflated without such tests.

    Authors: We acknowledge the concern that ML methods can potentially exploit simulation-specific features. Our primary results use held-out events from the identical simulation set to provide a matched-conditions benchmark, which is a prerequisite for any subsequent robustness study. In the revised manuscript we have added quantitative tests evaluating the trained model on simulations with perturbed scattering and absorption lengths (±10% and ±20%) as well as modified hole-ice models. These tests show that the relative resolution gains over the B-spline likelihood remain largely intact under moderate variations. Full retraining across an ensemble of ice models is computationally prohibitive at present and is identified as future work; we have updated the discussion section to emphasize this limitation. revision: partial

standing simulated objections not resolved
  • Direct real-data validation of angular resolution, as true neutrino directions are unknown for observed events.

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper trains a transformer encoder plus spherical normalizing flow model on Monte Carlo simulations to perform neural posterior estimation for neutrino directions, then evaluates median angular resolution on held-out test simulations against B-spline likelihood baselines. The reported factors (1.3× for throughgoing tracks, 1.7× for showers, 2.5× for starting tracks at 100 TeV) are direct empirical measurements from this train/test split; they do not reduce by any equation or definition in the paper to quantities defined in terms of the model's own fitted parameters or self-referential predictions. No self-definitional steps, fitted-input-as-prediction patterns, load-bearing self-citations, or ansatzes smuggled via prior work appear in the architecture description or performance claims. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that simulated events faithfully represent real detector behavior and that the learned posterior accurately captures directional uncertainty without overfitting to simulation artifacts.

free parameters (2)
  • Transformer hyperparameters
    Number of layers, heads, embedding dimension, and other architectural choices selected during development to optimize test performance.
  • Normalizing flow spline and transformation parameters
    Parameters of the rational-quadratic splines, scale, and rotation components learned from data during training.
axioms (1)
  • domain assumption Simulated IceCube events accurately model real detector response and ice properties
    All performance metrics are derived from simulation; no real-data validation is mentioned in the abstract.

pith-pipeline@v0.9.0 · 7896 in / 1406 out tokens · 98503 ms · 2026-05-10T01:11:36.699771+00:00 · methodology

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

58 extracted references · 1 canonical work pages

  1. [1]

    The IceCube Neutrino Observatory: instrumentation and online systems.Journal of Instrumentation, 12(03):P03012, 2017

    The IceCube Collaboration. The IceCube Neutrino Observatory: instrumentation and online systems.Journal of Instrumentation, 12(03):P03012, 2017

  2. [2]

    Evidence for neutrino emission from the nearby active galaxy NGC 1068.Science, 378(6619):538–543, 2022

    The IceCube Collaboration. Evidence for neutrino emission from the nearby active galaxy NGC 1068.Science, 378(6619):538–543, 2022

  3. [3]

    Observation of high-energy neutrinos from the Galactic plane

    The IceCube Collaboration. Observation of high-energy neutrinos from the Galactic plane. Science, 380(6652):1338–1343, 2023

  4. [4]

    P. A. Cherenkov. Visible luminescence of pure liquids under the influence ofγ-radiation.Dokl. Akad. Nauk SSSR, 2(8):451–454, 1934

  5. [5]

    Observation of High-Energy Astrophysical Neutrinos in Three Years of IceCube Data.Phys

    The IceCube Collaboration. Observation of High-Energy Astrophysical Neutrinos in Three Years of IceCube Data.Phys. Rev. Lett., 113:101101, 2014

  6. [6]

    Combining Maximum-Likelihood with Deep Learning for Event Reconstruction in IceCube

    The IceCube Collaboration. Combining Maximum-Likelihood with Deep Learning for Event Reconstruction in IceCube. InProceedings of 37th International Cosmic Ray Conference — PoS(ICRC2021), volume 395, page 1065, 2021

  7. [7]

    A convolutional neural network based cascade reconstruction for the IceCube Neutrino Observatory.Journal of Instrumentation, 16(07):P07041, 2021

    The IceCube Collaboration. A convolutional neural network based cascade reconstruction for the IceCube Neutrino Observatory.Journal of Instrumentation, 16(07):P07041, 2021

  8. [8]

    Improved modeling of in-ice particle showers for IceCube event reconstruction.Journal of Instrumentation, 19(06):P06026, 2024

    The IceCube Collaboration. Improved modeling of in-ice particle showers for IceCube event reconstruction.Journal of Instrumentation, 19(06):P06026, 2024

  9. [9]

    PhD thesis, University of Mainz, 2014

    Kai Schatto.Stacked searches for high-energy neutrinos from blazars with IceCube. PhD thesis, University of Mainz, 2014

  10. [10]

    Neutrino emission from the direction of the blazar TXS 0506+056 prior to the IceCube-170922A alert.Science, 361(6398):147–151, 2018

    The IceCube Collaboration. Neutrino emission from the direction of the blazar TXS 0506+056 prior to the IceCube-170922A alert.Science, 361(6398):147–151, 2018

  11. [11]

    A muon-track reconstruction exploiting stochastic losses for large-scale Cherenkov detectors.Journal of Instrumentation, 16(08):P08034, 2021

    The IceCube Collaboration. A muon-track reconstruction exploiting stochastic losses for large-scale Cherenkov detectors.Journal of Instrumentation, 16(08):P08034, 2021

  12. [12]

    Characterization of the astrophysical diffuse neutrino flux using starting track events in IceCube.Phys

    The IceCube Collaboration. Characterization of the astrophysical diffuse neutrino flux using starting track events in IceCube.Phys. Rev. D, 110:022001, 2024

  13. [13]

    State of the Ice Model in the IceCube Observatory

    The IceCube Collaboration. State of the Ice Model in the IceCube Observatory. In Proceedings of 39th International Cosmic Ray Conference — PoS(ICRC2025), volume 501, page 1013, 2025

  14. [14]

    Penalized splines for smooth representation of high-dimensional monte carlo datasets.Computer Physics Communications, 184(9):2214–2220, 2013

    Nathan Whitehorn, Jakob van Santen, and Sven Lafebre. Penalized splines for smooth representation of high-dimensional monte carlo datasets.Computer Physics Communications, 184(9):2214–2220, 2013

  15. [15]

    Energy reconstruction methods in the IceCube neutrino telescope

    The IceCube Collaboration. Energy reconstruction methods in the IceCube neutrino telescope. Journal of Instrumentation, 9(03):P03009, 2014

  16. [16]

    The IceCube realtime alert system.Astroparticle Physics, 92:30–41, 2017

    The IceCube Collaboration. The IceCube realtime alert system.Astroparticle Physics, 92:30–41, 2017

  17. [17]

    Fastε-free inference of simulation models with Bayesian conditional density estimation

    George Papamakarios and Iain Murray. Fastε-free inference of simulation models with Bayesian conditional density estimation. InProceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 1036–1044, Red Hook, NY, USA,

  18. [18]

    Curran Associates Inc

  19. [19]

    The frontier of simulation-based inference.Proceedings of the National Academy of Sciences, 117(48):30055–30062, 2020

    Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The frontier of simulation-based inference.Proceedings of the National Academy of Sciences, 117(48):30055–30062, 2020

  20. [20]

    Conditional normalizing flows for IceCube event reconstruction

    The IceCube Collaboration. Conditional normalizing flows for IceCube event reconstruction. InProceedings of 38th International Cosmic Ray Conference — PoS(ICRC2023), volume 444, page 1003, 2023

  21. [21]

    Graph Neural Networks for low-energy event classification & reconstruction in IceCube.Journal of Instrumentation, 17(11):P11003, 2022

    The IceCube Collaboration. Graph Neural Networks for low-energy event classification & reconstruction in IceCube.Journal of Instrumentation, 17(11):P11003, 2022. 38

  22. [22]

    Dynamic Graph CNN for Learning on Point Clouds

    Yue Wang, Yongbin Sun, Ziwei Liu, et al. Dynamic Graph CNN for Learning on Point Clouds. ACM Trans. Graph., 38(5), 2019

  23. [23]

    Attention is All you Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, et al. Attention is All you Need. InAdvances in Neural Information Processing Systems, volume 30, 2017

  24. [24]

    Normalizing Flows on Tori and Spheres

    Danilo Jimenez Rezende, George Papamakarios, Sebastien Racaniere, et al. Normalizing Flows on Tori and Spheres. InProceedings of the 37th International Conference on Machine Learning, volume 119, pages 8083–8092. PMLR, 2020

  25. [25]

    Neural Spline Flows

    Conor Durkan, Artur Bekasov, Iain Murray, et al. Neural Spline Flows. InAdvances in Neural Information Processing Systems, volume 32, pages 7511 – 7522, 2019

  26. [26]

    The Evolution of Markov Chain Monte Carlo Methods.The American Mathematical Monthly, 117(5):383–413, 2010

    Matthew Richey. The Evolution of Markov Chain Monte Carlo Methods.The American Mathematical Monthly, 117(5):383–413, 2010

  27. [27]

    Jordan, Zoubin Ghahramani, Tommi S

    Michael I. Jordan, Zoubin Ghahramani, Tommi S. Jaakkola, et al. An Introduction to Variational Methods for Graphical Models.Machine Learning, 37(2):183–233, 1999

  28. [28]

    Kullback and R

    S. Kullback and R. A. Leibler. On information and sufficiency.The Annals of Mathematical Statistics, 22(1):79–86, 1951

  29. [29]

    Thorsten Glüsenkamp. Unifying supervised learning and VAEs: coverage, systematics and goodness-of-fit in normalizing-flow based neural network models for astro-particle reconstructions.The European Physical Journal C, 84:163, 2024

  30. [30]

    Dispersion on a Sphere.Proceedings of the Royal Society of London Series A, 217(1130):295–305, 1953

    Ronald Fisher. Dispersion on a Sphere.Proceedings of the Royal Society of London Series A, 217(1130):295–305, 1953

  31. [31]

    Numerically stable sampling of the von Mises Fisher distribution onS2 (and other tricks).EPFL research report, 2012

    Wenzel Jakob. Numerically stable sampling of the von Mises Fisher distribution onS2 (and other tricks).EPFL research report, 2012

  32. [32]

    Higham.Accuracy and Stability of Numerical Algorithms

    Nicholas J. Higham.Accuracy and Stability of Numerical Algorithms. Society for Industrial and Applied Mathematics, second edition, 2002

  33. [33]

    Glüsenkamp

    T. Glüsenkamp. Jammy Flows.https://github.com/thoglu/jammy_flows

  34. [34]

    K. M. Górski, E. Hivon, A. J. Banday, et al. HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere.The Astrophysical Journal, 622(2):759, 2005

  35. [35]

    MOC - HEALPix Multi-Order Coverage map Version 1.0.International Virtual Observatory Alliance, 2014

    Pierre Fernique, Thomas Boch, Tom Donaldson, et al. MOC - HEALPix Multi-Order Coverage map Version 1.0.International Virtual Observatory Alliance, 2014

  36. [36]

    Martinez-Castellanos, Leo P

    I. Martinez-Castellanos, Leo P. Singer, E. Burns, et al. Multiresolution HEALPix Maps for Multiwavelength and Multimessenger Astronomy.The Astronomical Journal, 163(6):259, 2022

  37. [37]

    Lawson and Richard J

    Charles L. Lawson and Richard J. Hanson.Solving least squares problems. Englewood Cliffs, NJ: Prentice-Hall, 1974

  38. [38]

    On layer normalization in the transformer architecture

    Ruibin Xiong, Yunchang Yang, Di He, et al. On layer normalization in the transformer architecture. InProceedings of the 37th International Conference on Machine Learning, ICML’20. JMLR.org, 2020

  39. [39]

    Swart, and George Obaido

    Ibomoiye Domor Mienye, Theo G. Swart, and George Obaido. Recurrent Neural Networks: A Comprehensive Review of Architectures, Variants, and Applications.Information, 15(9), 2024

  40. [40]

    Learning phrase representations using RNN encoder–decoder for statistical machine translation

    Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, et al. Learning phrase representations using RNN encoder–decoder for statistical machine translation. InProceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724–1734, Doha, Qatar, 2014. Association for Computational Linguistics

  41. [41]

    A model independent parametrization of the optical properties of the refrozen IceCube drill holes

    The IceCube Collaboration. A model independent parametrization of the optical properties of the refrozen IceCube drill holes. InProceedings of 38th International Cosmic Ray Conference — PoS(ICRC2023), volume 444, page 1034, 2023

  42. [42]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In Proceedings of the 3rd International Conference on Learning Representations, 2015. 39

  43. [43]

    FlashAttention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR), 2024

  44. [44]

    SGDR: Stochastic Gradient Descent with Warm Restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradient Descent with Warm Restarts. InInternational Conference on Learning Representations, 2017

  45. [45]

    A Variational Analysis of Stochastic Gradient Algorithms

    Stephan Mandt, Matthew Hoffman, and David Blei. A Variational Analysis of Stochastic Gradient Algorithms. InProceedings of The 33rd International Conference on Machine Learning, volume 48 ofProceedings of Machine Learning Research, pages 354–363, New York, New York, USA, 2016. PMLR

  46. [46]

    Averaging Weights Leads to Wider Optima and Better Generalization

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, et al. Averaging Weights Leads to Wider Optima and Better Generalization. InProceedings of the Thirty-Fourth Conference on Uncertainty in Artificial Intelligence, UAI 2018, Monterey, California, USA, August 6-10, 2018, pages 876–885. AUAI Press, 2018

  47. [47]

    Self-Attention with Relative Position Representations

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-Attention with Relative Position Representations. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 464–468, New Orleans, Louisiana, 2018. Association for Computational Linguistics

  48. [48]

    IceCube – Neutrinos in Deep Ice

    Habib Bukhari, Dipam Chakraborty, Philipp Eller, et al. IceCube – Neutrinos in Deep Ice. The European Physical Journal C, 84(6):646, 2024

  49. [49]

    ResiDual: Transformer with Dual Residual Connections.arXiv:2304.14802, 2024

    Shufang Xie, Huishuai Zhang, Junliang Guo, et al. ResiDual: Transformer with Dual Residual Connections.arXiv:2304.14802, 2024

  50. [50]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInternational Conference on Learning Representations, 2021

  51. [51]

    Estimating the angular resolution of tracks in neutrino telescopes based on a likelihood analysis.Astroparticle Physics, 25(3):220–225, 2006

    Till Neunhöffer. Estimating the angular resolution of tracks in neutrino telescopes based on a likelihood analysis.Astroparticle Physics, 25(3):220–225, 2006

  52. [52]

    C. A. Argüelles, A. Schneider, and T. Yuan. A binned likelihood for stochastic models. Journal of High Energy Physics, 2019(6):30, 2019

  53. [53]

    Thomas K. Gaisser. Spectrum of cosmic-ray nucleons, kaon production, and the atmospheric muon charge ratio.Astroparticle Physics, 35(12):801–806, 2012

  54. [54]

    Gaisser, and Todor Stanev

    Anatoli Fedynitch, Felix Riehn, Ralph Engel, Thomas K. Gaisser, and Todor Stanev. Hadronic interaction model sibyll2.3cand inclusive lepton fluxes.Phys. Rev. D, 100:103018, 2019

  55. [55]

    Improved measurements of the TeV-PeV extragalactic neutrino spectrum from joint analyses of IceCube tracks and cascades.Phys

    The IceCube Collaboration. Improved measurements of the TeV-PeV extragalactic neutrino spectrum from joint analyses of IceCube tracks and cascades.Phys. Rev. D, 113:062002, 2026

  56. [56]

    The IceCube Upgrade: status and prospects for advances with GeV neutrinos

    The IceCube Collaboration. The IceCube Upgrade: status and prospects for advances with GeV neutrinos. InProceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024), volume 476, page 149, 2024

  57. [57]

    IceCube-Gen2: the window to the extreme Universe.Journal of Physics G: Nuclear and Particle Physics, 48(6):060501, 2021

    The IceCube Collaboration. IceCube-Gen2: the window to the extreme Universe.Journal of Physics G: Nuclear and Particle Physics, 48(6):060501, 2021

  58. [58]

    J. A. Gregory and R. Delbourgo. Piecewise Rational Quadratic Interpolation to Monotonic Data.IMA Journal of Numerical Analysis, 2(2):123–130, 1982. 40