acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Unified Neutrino Event Reconstruction via Task-Aware Retrieval and Geometry-Grounded Decoding

Abstract

Neutrino telescopes infer multiple latent physical quantities from the same sparse, irregular spatio-temporal pulse array, yet heterogeneous reconstruction targets depend on fundamentally distinct observational evidence. We present PRISM, a unified reconstruction framework that addresses this challenge through a three-level decoupling: shared event representation, task-conditioned information retrieval, and targeted structural specialization. PRISM employs a Transport Skeletonizer (TSK) to maintain a compact event-level latent memory while preserving fine-grained hit representations, and learns task-aware queries to independently extract target-dependent features from the combined event-level and hit-level representations. For interaction-vertex localization, a geometry-aware coordinate anchor grounds the decoder with an explicit sensor-coordinate reference prior to unconstrained residual refinement. Across the formal NuBench benchmark, PRISM offers a favorable accuracy–efficiency trade-off: it achieves competitive performance with published per-cell reference results in selected energy and topology regimes, while requiring approximately fewer parameters and providing – higher inference throughput than sequential multi-model execution. Rigorous equal-compute controls attribute a 32.65% reduction in median vertex error to coordinate anchoring, while ablations and controlled perturbation tests demonstrate task-dependent observation stability rather than universal robustness. Directional reconstruction, absolute vertex precision, and low-energy events remain key open challenges. Overall, our findings support a unified reconstruction strategy that shares expensive spatio-temporal representations, retrieves target-specific evidence, and introduces physically grounded specialization where uniform sharing is insufficient.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multi-Task Learning, AI for Science, Task-Aware Decoding, Point Cloud Learning, Neutrino Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_V1dcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a unified multi-task framework for neutrino telescope event reconstruction that combines a Transport Skeletonizer (TSK), task-aware queries, and a geometry-aware coordinate anchor. The framework aims to jointly reconstruct five event-level properties while sharing the underlying unified representation. However, the experimental evidence is insufficient to demonstrate a competitive accuracy–efficiency trade-off compared with existing neutrino reconstruction methods, which is claimed as the main contribution of the paper.

Critical strengths

Exploring a unified representation for neutrino telescope event reconstruction is a meaningful direction. The proposed architecture is a reasonable engineering choice for balancing local and global event information at a manageable computational cost. In particular, the persistent skeleton tokens allow the model to capture event-level structures, such as extended particle tracks spanning multiple local patches, without calculating the full attention across all hits. Task-aware queries are also a reasonable design choice, as different reconstruction tasks may rely on different features of the event.

The ablation studies of the main components are informative, especially the equal-compute comparison for the coordinate anchor. However, some important comparisons are still missing, as discussed in the Weaknesses section.

Critical weaknesses

Minor issue: In figure 1 task (e) is introduced as a classification task, which is not consistent with other parts of the paper. If it’s a typo, please fix it.

First of all, the writing of this paper needs to be improved a lot. The current introduction and problem setup is not very clear for general AI4S readers. For example, the authors should explain the following jargons/items more clearly:

  • Detector-string identity: This might be commonly used in the field of neutrino telescope, but not a commonly known jargon
  • The meaning of each variables in y = {y_{topo}, E, d, v, y_{vis}}
  • u_i is never defined
  • phi_{in} and e are introduced as embeddings, but their implementation is not explained

Second, the methodological novelty of the Transport Skeletonizer (TSK) is unclear. Its core mechanism is similar to Recurrent Interface Networks (RIN; Jabri et al., ICML 2023), which already employ learnable latent tokens to read information from input tokens and write it back through cross-attention. RIN also preserves input-level representations and performs multiple rounds of read-write operations with persistent latent states. Furthermore, FIT (Chen & Li, 2023) already combines local attention with global latent features, which matches the overall architecture adopted by PRISM. This paper does not sufficiently distinguish TSK from these prior works. As an application-track paper, developing an original algorithm is not necessarily required. However, the authors should properly acknowledge and discuss existing methods rather than re-inventing established mechanisms with a new name and claiming them as novel contributions.

Furthermore, the ablation studies are insufficient to support some of the main architectural claims. One of the central motivations of TSK is to preserve fine-grained hit-level representations, and combine the information with event-level skeleton tokens. However, the authors do not compare the model performance in those 3 cases:

  1. Using only hit-level representations
  2. Using only skeleton representations
  3. their combination in the final task-aware retrieval stage (main result)

Without this comparison, the experiments cannot justify the motivation and effectiveness to keep both representations. A single concatenation can not be justified as contribution without clear improvement demonstrated. Besides, the TSK ablation only compares the proposed global information exchange mechanism against removing it entirely. This demonstrates the benefit of introducing global information exchange compared with the local attention only case, but does not establish the advantage of TSK over existing alternatives such as RIN or FIT style mechanisms.

The inference efficiency comparison in Table 1(c) is fundamentally uninformative. It is common knowledge that running five copies of the same model sequentially will require approximately five times the computational cost of running a single copy. Although sharing event representations is a reasonable approach to reducing redundant computation, the current comparison provides no meaningful evidence of the computational efficiency of PRISM relative to existing alternatives. If the authors want to justify the computational efficiency of the proposed framework, they should at least compare against:

  1. Existing single-task baselines: The inference cost of each of the existing reconstruction models reported in Table 1(a), including their parameter counts, FLOPs, etc.
  2. A conventional unified multi-task Transformer: An independently trained model with a shared encoder and five task-specific output heads, but without TSK or task-aware queries. This would provide a more meaningful baseline for determining whether the proposed architectural components offer any computational advantage beyond standard representation sharing.The shared pooled multi-task baseline mentioned in Table2 could serve as such a baseline, but neither the detailed structure nor the computation cost is reported.

Finally, the most serious weakness of this paper is that the proposed model matches or outperforms the best published reference results in only 12 out of 136 evaluated metric cells across the four detector configurations. In particular, PRISM fails to reach the best published performance in any of the direction or vertex reconstruction metrics. Although the reference results are taken from different models rather than a single baseline, these results raise critical concerns about the competitiveness of the proposed framework. Given that the primary motivation of this work is to develop a unified framework for neutrino telescope event reconstruction, the authors need to justify the practical effectiveness and scientific impact of the proposed model, especially considering its limited reconstruction accuracy compared with existing methods. Without convincing evidence, for example, the computational savings can compensate for the loss in reconstruction accuracy, it is difficult to establish the practical value of the proposed framework.

Questions for authors
  1. Given that PRISM only matches or outperforms the best published results in 12 out of 136 metric cells, what is the practical scientific value of this framework? In what practical scenarios would PRISM be preferred over the existing reconstruction models?
  2. Can the authors provide a meaningful inference efficiency comparison against the existing single-task baselines and a conventional shared-encoder multi-task model? How much computational cost can PRISM actually save at comparable reconstruction accuracy, and whether the amount is substantial enough to be considered as a critical factor for similar tasks?
  3. What are the main methodological differences between TSK and existing architectures such as RIN and FIT?
  4. Can the authors compare the performance of using only hit representations, only skeleton representations, and both in the final task-aware retrieval stage?
4 comments
AuthorsOpenReview ↗

Response to Reviewer V1dc (Part 1/2)

1. Practical and Scientific Value (Q1/W5):

Practical value. Neutrino-event analyses often require multiple physical observables from the same detector event. PRISM provides all five reconstruction outputs through a single shared forward pass, avoiding repeated event encoding across independently trained task-specific networks. On Flower-XL, it achieves better reconstruction accuracy than the evaluated small-specialist controls while reducing five-output FP32 inference latency by 3.98× (8.420 vs. 33.497 ms/event). This offers a practical alternative for multi-output reconstruction pipelines where inference throughput and shared computation are important.

Scientific value. Beyond computational efficiency, PRISM provides insight into how heterogeneous physical reconstruction tasks should share and retrieve information. Our task-subset experiments reveal non-monotonic transfer between objectives, while hit/skeleton retrieval ablations show that different targets benefit from different representations. These findings motivate PRISM's three-level design: shared event representation, task-aware information retrieval, and targeted geometric specialization. Thus, PRISM contributes a framework for investigating how shared physical evidence can support distinct reconstruction objectives, rather than merely combining five prediction heads.

Benchmark positioning. On the published NuBench reference envelope, PRISM matches or exceeds the best reported results in 12/136 metric cells, primarily in topology and energy. Strong single-task references retain advantages in several direction and vertex regimes. These results distinguish PRISM's contribution from single-task SOTA optimization: its focus is jointly reconstructing heterogeneous observables through shared computation while maintaining competitive performance in selected reconstruction tasks.

2. Multi-Task Baselines and Inference Efficiency (Q2/W4):

The submitted five-copy comparison in Table 1(c) measures duplicated computation. We additionally evaluated a pooled multi-task baseline, independent specialists, and retained external implementations.

Pooled baseline. Submitted Table 2 uses the same six-layer local encoder without TSK/task queries, with masked mean pooling and five linear heads (seeds 42/43).

Table G: Original seed-42 checkpoints; Batch-32 FP32 timing on RTX A6000, PyTorch 2.5.1/CUDA 12.4 (128 events, 50 warm-ups, five repeats, ≥200 synchronized calls; transfers excluded). GFLOPs count matrix operations only.

ModelParams(M)Ewidth(%)↓Dir(°)↓Vertex(m)↓Latency(ms/event)↓MatmulGFLOPs/event
Base-PRISM6.02565.8358.17617.2297.00623.165
Pooled2.90892.94210.91733.2085.79415.701

Compared with conventional pooled multi-task reconstruction, PRISM substantially improves energy, direction, and vertex accuracy, at a higher parameter and inference cost. The comparison quantifies the accuracy–efficiency trade-off between the evaluated shared-encoder architectures.

Independent specialists. Five separately trained small models were pruned of unused branches; outputs were unchanged on 128 checked events at Batch 1/32 in FP32/FP16.

Table H: Batch-32 five-output inference after matched 20k-update vertex-only Anchor refinement, with sequential specialists on one GPU. FP16 uses autocast.

SettingDeployedParams(M)FP32(ms/event)↓FP16(ms/event)↓
Joint-PRISM+Anchor7.8458.4203.416
Five-pruned-small-ST6.11733.49716.781

Joint PRISM achieves 3.98× FP32 and 4.91× FP16 inference speedups over sequential specialists. The FP32 results also favor joint reconstruction across all five tasks after Anchor refinement. The comparison uses 7.845M vs. 6.117M deployed parameters, FP16 results concern latency only.

External implementations. Under a separate RTX A6000/PyTorch 2.1/CUDA 11.8 FP32 protocol, Batch-32 forward costs (parameters; ms/event) were PRISM (7.845M; 5.459), DeepIce (115.351M; 64.110), DynEdge (1.358M; 0.248), and ParticleNet (0.326M; 0.239). These measurements illustrate the computational diversity of existing reconstruction architectures: lightweight GNNs have substantially lower single-model inference latency, whereas larger architectures such as DeepIce incur higher costs. Table H separately evaluates the five-output inference setting with independently trained specialists, providing direct evidence of PRISM's computational sharing benefit.

AuthorsOpenReview ↗

Response to Reviewer V1dc (Part 2/2)

3. Relationship to RIN and FIT (Q3/W2):

Methodological relationship. RIN (Jabri et al., 2023) and FIT (Chen and Li, 2023) provide important precedents for latent-mediated information exchange. PRISM builds on this established paradigm and adapts it to sparse, geometry-structured detector events. Its design combines Morton-ordered local encoding, persistent event-level skeleton tokens, gated hit-level feedback, task-conditioned retrieval, and coordinate-aware vertex refinement. We will clarify this relationship and cite both works in the revised manuscript.

Controlled comparisons. We evaluated RIN-style (latent self-attention/FFN between read/write steps) and FIT-style (one latent per 64-hit group with local/global exchange) adaptations under matched Flower-XL training exposure and task losses. Both retain PRISM's surrounding architecture, allowing comparison of the exchange mechanisms within a common reconstruction framework. These are adaptations rather than full reproductions of the original architectures.

Table I: Full Flower-XL test, seed 42, before Anchor; parameters/FLOPs are not matched.

ExchangeParams(M)TopoAUC↑Ewidth(%)↓Dir(°)↓Vertex(m)↓Visible-y↓
PRISM-TSK7.8060.905561.8327.76619.0820.0891
RIN-style9.1410.904062.0357.71221.1150.0905
FIT-style9.1400.870096.08912.86026.9030.1263

Experimental comparison. Under matched training exposure, PRISM-TSK achieves performance broadly comparable to the RIN-style adaptation on topology, energy, direction, and visible-y, while reducing vertex error from 21.115 m to 19.082 m with approximately 14.6% fewer parameters. It also outperforms the evaluated FIT-style adaptation across the reported tasks. These results characterize the behavior of the three exchange designs in the considered reconstruction setting.

4. Hit, Skeleton, and Combined Retrieval (Q4/W3):

We compared Hits-only, Skeleton-only, and Both under matched Flower-XL exposure. All retain encoder TSK read/write exchanges; only the final task-query memory and shared summary differ. The experiment isolates the contribution of final task-conditioned retrieval while retaining the same TSK read/write encoder across variants.

Table J: Full Flower-XL test (2,980,023 events), seed 42, three epochs, effective batch 96, before Anchor; independent task decoders. Both is the supplementary joint control, not submitted Table 2 Base.

RetrievalTopoAUC↑Ewidth(%)↓Dir(°)↓Vertex(m)↓Visible-y↓
Hits-only0.896272.6448.63818.8710.1014
Skeleton-only0.901674.1238.16715.2980.0967
Both0.905561.8327.76619.0820.0891

Combined retrieval benefits four tasks, but Skeleton-only has lower vertex error. Paired event-bootstrap 95% CIs for Both − Hits are [−10.910,−10.695] pp (energy width), [−0.882,−0.859]° (direction), and [+0.195,+0.228] m (vertex). Both − Skeleton vertex is +3.784 m (95% CI [3.770,3.800] m). Thus, gains are task-dependent; the separate equal-compute Anchor ablation supports vertex-specific geometric specialization. Hits-only did not restore the original worker RNG state; bootstrap CIs reflect event-sampling uncertainty for fixed checkpoints, not seed variability.

5. Task Definitions and Input Notation (W1)

Figure 1(e). Visible inelasticity is continuous regression, not classification: a bounded scalar output trained with Smooth-L1 loss and evaluated on CC events. Reported metrics are unaffected; we will relabel it "Visible Inelasticity Regression."

Targets. The five targets are ytopoy_{\mathrm{topo}} (binary topology label), EE (neutrino energy), d\mathbf{d} (3D incoming unit direction), v\mathbf{v} (3D interaction vertex), and yvisy_{\mathrm{vis}} (continuous visible inelasticity). Topology labels do not imply universal CC/track equivalence.

Inputs. String identity sis_i denotes the sensor string associated with pulse ii, containing multiple optical sensors, and uses a categorical embedding. The numerical input is

ui=[xi/100,yi/100,zi/1000,ti/105,log⁡10(max⁡(qi,0)+1)]\mathbf{u}_i = [x_i/100, y_i/100, z_i/1000, t_i/10^5, \log_{10}(\max(q_i,0)+1)]

(coordinates in metres; time in native units). For Flower-XL, ϕin\phi_{\mathrm{in}} is Linear(5,192) → LayerNorm(192) → GELU, and e(si)e(s_i) is Embedding(1211,192); their outputs are added. We will clarify these definitions and correct Figure 1(e) in the revised manuscript.

Thank you for the detailed response and additional experiments.

The new hit/skeleton ablations adequately address Q4. The discussion and comparisons with RIN and FIT also address much of Q3. However, Q2 and my main concern about practical scientific value remain insufficiently addressed.

  1. Accuracy: PRISM matches or exceeds the best published results in only 12/136 metric cells. This substantial performance gap remains a major concern.

  2. Efficiency: The reported 3.98× speedup is measured against five small specialist models constructed by the authors, rather than existing competitive reconstruction methods as I suggested. The external baseline comparisons report inference costs without corresponding accuracy comparisons. Therefore, the experiments still do not answer whether PRISM offers a competitive accuracy–efficiency trade-off against existing methods.

I therefore maintain my original rating

AuthorsOpenReview ↗

Follow-up response

Thank you for the follow-up and for clarifying the remaining concerns.

External specialist accuracy–efficiency comparison. To address Q2, we have completed additional comparisons with official NuBench specialist architectures, including DynEdge, GRIT, ParticleNet, and DeepIce. Unlike our earlier PRISM-based single-task controls, these are independently designed reconstruction methods. We evaluate reconstruction accuracy on the same 2,980,023 Flower-XL test events and measure inference costs on an RTX A6000 (FP32, TF32 disabled, batch 32).

SettingDir (°) ↓Vertex (m) ↓Low-E width (%) ↓Topo AUC ↑Visible-y ↓Latency (ms/event) ↓
Joint PRISM8.05410.196123.250.90270.09135.814
5 DynEdge5.0186.744208.390.88870.094213.682
4 GRIT + DynEdge-D5.01833.348119.720.89940.067529.121
Fixed mixed specialists*3.4786.744116.570.89940.0675103.008

The fixed mixed portfolio combines DeepIce (direction), DynEdge (vertex), ParticleNet (energy), and GRIT (topology and visible-y). Accuracy uses full-test released/frozen predictions. Latency includes native preprocessing, graph construction, data transfer, and forward inference under a common serial CPU protocol.

Accuracy–efficiency trade-off. PRISM achieves lower serial five-output pipeline latency, while the external specialists provide stronger reconstruction on several objectives, particularly direction and vertex. These results characterize the task-dependent accuracy–cost trade-offs between unified reconstruction and dedicated specialist approaches. The corresponding comparisons are presented in Appendix G, Figure 4.

Scientific contribution. Performance gaps to published best results remain, particularly for direction and vertex reconstruction. Beyond computational efficiency, PRISM investigates a central question in unified scientific reconstruction: how to share event-level information while preserving task-specific evidence for heterogeneous physical targets. Our task-subset experiments reveal non-monotonic transfer between objectives, while hit/skeleton retrieval ablations demonstrate differing information requirements. These findings support PRISM's task-aware retrieval and structural specialization, while offering empirical insights into the interplay between shared event representations and task-specific information requirements in unified event reconstruction.

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

In this paper, a neutrino-event reconstruction framework, named PRISM, is proposed, which is established in a unified manner, simultaneously reconstructing five aspects of critical neutrino events. It comprises a shared local encoder, a transport skeletonizer that preserves hit-level representations, task-specific query decoders, and a coordinate anchor with residual refinement for vertex localization. Several metrics are evaluated on public benchmarks and reported to support the paper's central claim about performance improvements and reduction of inference cost. My main concerns are the interpretation of certain reported numbers that reflect potential inconsistency between the documented methods and the reported values, which could suggest the invalidity of the numerical results. I am also not convinced by the accuracy-efficiency trade-off.

Critical strengths
  • Unified reconstruction is a particular advanced topic in the context of particle physics. The distinction between shared representations and task-specific heads is a natural design following the big trend of foundation models. These parts are clearly well-motivated.
  • The evaluation follows the commonly used NuBench and its publicly available procedure, which means the results have high reproducibility. The central claims, if well supported, are certainly great progress in this particular field. Several experiments are done with care, nevertheless.
  • I appreciate the paper's honesty towards negative results and limitations.
Critical weaknesses

1. The low-energy results need clarification.

I appreciate the authors list key results as early as Table 1. However, when trying to trace further in the appendix, I find the Table 4 numbers hard to swallow. In particular, it reports low-energy signed bias/width of +189.56%/369.62% on Flower-L and +333.19%/901.26% on Cluster.

These numbers are surprisingly large and make me try to scrutinize further to understand. Then I found that Equation 17 defines the residual as 100(1 - E_reco/E_true). Under this definition, nonnegative reconstructed energies cannot produce residuals above 100%. If both the stated convention and table signs are correct, these median biases imply negative reconstructed energies for at least half the events in the affected bins. I don't know how this could happen, but it's easy to fix at the model level...

This could reflect the prediction parameterization, target conversion, or a reporting/convention inconsistency. I couldn't figure it out from the manuscript, and would like to see a clarification. It could be just one or two numbers being off, or the whole evaluation becomes questionable.

2. The practical accuracy-efficiency trade-off is not measured in a good way.

When discussing the reduction of the inference cost, the current paper makes use of 5 passes of the same model. While I appreciate that the manuscript clearly identifies the latter as a storage/compute proxy. Nevertheless, the approximately fivefold saving is largely a consequence of executing one model instead of five rather than a trade-off. Clearly, an expert model that focuses on a certain aspect might not need to be as large as a unified foundation model.

This matters especially to me since I noticed that direction and vertex gaps are substantial. For example, on high-energy Cluster CC events, direction error is 6.49 degrees versus DeepIce's 1.22 degrees; high-energy vertex error is 4.84 m versus DynEdge's 1.29 m (Tables 5 and 7). I see off, but not necessarily traded for inference improvements. I would suggest listing these compared models' inference costs, etc.

Questions for authors
  1. Can you reconcile Equation 17 with Table 4's positive biases above 100%, specifying the energy target/output transformation?
  2. Can you provide matched single-task direction/vertex results with measured inference cost, or narrow the accuracy-efficiency claim?
  3. Is it possible to keep a more concrete metrics table in the main text? I find the current reading flow a bit twisted.
  4. Can you specify training batch size, complete optimization exposure, timing hardware/protocol, and the anchor validation composite, and report uncertainty for the central comparisons?
3 comments
AuthorsOpenReview ↗

Response to Reviewer gVqA

1. Correcting the Energy-Residual Convention (Q1/W1):

We thank the reviewer for identifying the sign-convention mismatch in the evaluator. The model predicts log⁡10(E/GeV)\log_{10}(E/\mathrm{GeV}), which is converted to a strictly positive reconstructed energy using Ereco=10outputE_{\mathrm{reco}}=10^{\mathrm{output}} GeV. The evaluator used 100(Ereco/Etrue−1)100(E_{\mathrm{reco}}/E_{\mathrm{true}}-1), whereas Equation 17 defines 100(1−Ereco/Etrue)100(1-E_{\mathrm{reco}}/E_{\mathrm{true}}).

Auditing all four configurations reproduced 24/24 energy entries. Only the 12 signed-bias entries require sign reversal, all 12 widths and absolute-bias rankings remain unchanged. No nonpositive energies or clipping were found. Replays of 96 stratified events per configuration for Flower-L/Cluster matched saved log-energy predictions within 2.15×10⁻⁶/9.54×10⁻⁷.

The corrected low-energy biases are −28.81%, −189.56%, −85.27%, and −333.19% for Flower-XL, Flower-L, Hexagon, and Cluster, respectively. Model predictions and accuracy are unchanged. The corrected signs indicate substantial low-energy overestimation. We will correct PRISM bias signs while retaining published NuBench references.

2. Matched Specialist Comparisons and Efficiency Scope (Q2/W2):

The submitted five-copy comparison measures duplicated computation. We therefore trained independent direction and vertex single-task (ST) controls alongside a five-task joint model under matched Flower-XL conditions (seed 42, three epochs, effective batch 96). Both vertex models received 20k Anchor-refinement updates. These controls are separate from submitted models.

Table D: Official-test accuracy and Batch-32 FP32 latency. ST branches were pruned with output equivalence verified; parameters are deployed counts.

SettingDir(°)↓Vertex(m)↓Params(M)Latency(ms/event)↓
Direction-ST12.748—6.0247.850
Vertex-ST—15.3416.0247.849
Joint-initial7.76619.0827.8068.257
Vertex-ST+20k-Anchor—18.5456.0627.858
Joint+20k-Anchor7.76613.1507.8458.279

Joint training improves direction but initially has higher vertex error; after Anchor refinement its vertex error is 13.150 m, versus 18.545 m for refined ST. Single-output ST inference remains cheaper.

Under PyTorch 2.1/CUDA 11.8, Batch-32 latencies were 5.459/64.110/0.248 ms/event for PRISM/DeepIce/DynEdge (7.845M/115.351M/1.358M parameters). Different inputs/training configurations preclude controlled accuracy–efficiency rankings. On Cluster high-energy events, PRISM versus specialists yields 6.488°/1.22° (CC direction, DeepIce) and 4.840/1.29 m (vertex, DynEdge).

In a separate small-ST benchmark, PRISM achieves 8.420 vs. 33.497 ms/event, corresponding to a 3.98× FP32 speedup.

3. Concrete Reconstruction Metrics (Q3):

Table E: Original submitted checkpoints, with corrected energy-bias signs; supplementary ST/joint controls are separate. Low-E: 10–100 GeV; high-E: >1 TeV. Direction: high-E CC; vertex: all high-E; topology AUC: overall; visible-y median: all-energy CC. Width = P84−P16.

ConfigurationLow-E-bias/width(%)High-E-CC-Dir(°)↓High-E-Vertex(m)↓TopoAUC↑Visible-y-median↓
Flower-XL−28.81/123.253.587.690.90270.0913
Flower-L−189.56/369.623.773.810.93740.0865
Hexagon−85.27/219.283.115.380.93540.0806
Cluster−333.19/901.266.494.840.84560.1330

4. Training Protocol, Timing, and Uncertainty (Q4):

Training. AdamW (lr=3e-4, weight decay 0.05), gradient clipping 1.0, cosine decay with 3,750 warm-up steps. Anchor lr=3e-5.

Table F: Audited training budgets (I/A=Initial/Anchor).

ConfigurationBatch(I/A)Updates(I/A)Events(I/A)
Flower-XL96/96213,691/213,68320.52M/20.52M
Flower-L64/64952,582/952,51260.99M/60.99M
Hexagon192/64256,771/770,30649.32M/49.32M
Cluster72/72300,000/805,07421.61M/57.99M

Initial models use their last checkpoints. Anchor checkpoints are selected by the validation composite:

Topology accuracy − 0.02×Direction median (°) − 0.05×log10-energy MAE − 0.01×Vertex mean L2 (m) − 0.05×Visible-y MAE.

The composite uses validation-batch statistics; only vertex is refined. Exposures include post-selection updates; 20k controls use vertex validation loss.

Timing: Submitted Table 1(c) measures Batch-1/32 compute-only five-copy inference. Table D uses RTX A6000, PyTorch 2.5.1/CUDA 12.4, FP32 (TF32 off), Batch 32, 128 events, 50 warm-ups, five repeats, ≥200 synchronized calls; transfers/loading excluded. Protocols are not pooled.

Uncertainty: 1,000 paired event-bootstraps on Flower-XL (N=2,980,023) give 95% CIs for Joint − ST: energy width [−27.854,−27.572] pp, direction [−4.998,−4.967]°, initial vertex [+3.726,+3.759] m, refined vertex [−5.410,−5.380] m. Equal-compute Anchor reduces median vertex error 18.911→12.738 m (95% CI relative gain 32.05%–33.25%). CIs capture event sampling for fixed checkpoints, not seed variability.

Thank you for the detailed response, especially for clarifying the energy-residual convention and providing the additional single-task and efficiency comparisons. These help clarify the results.

The reported low-energy metrics nevertheless suggest that reconstruction in this regime remains challenging. More broadly, I therefore remain concerned about the reconstruction performance and whether the accuracy–efficiency tradeoff is sufficiently compelling for the proposed unified approach.

The additional single-task comparison is informative, but it seems to compare variants of the same architecture, right? It therefore does not by itself establish a computational advantage over dedicated specialist models at comparable reconstruction quality. I would find such a comparison important for assessing the practical benefit of the unified model.

Based on the above, I'd retain my current rating.

AuthorsOpenReview ↗

Follow-up response

Thank you for the follow-up and for emphasizing the importance of dedicated specialist comparisons.

External specialist comparison. Our previous single-task controls were independently trained variants of PRISM. We have now extended the evaluation to official NuBench architectures, including DynEdge, GRIT, ParticleNet, and DeepIce. Reconstruction accuracy is evaluated on the same 2,980,023 Flower-XL test events. Inference costs are measured on an RTX A6000 (FP32, TF32 disabled, batch 32).

Five-output accuracy–efficiency comparison. We evaluated three independent specialist portfolios against joint PRISM. Under a common serial preprocessing protocol, the measured costs include native preprocessing, graph construction, data transfer, and forward inference.

SettingDir(°)↓Vertex(m)↓Low-E width(%)↓TopoAUC↑Visible-y↓Latency(ms/event)↓
Joint PRISM8.05410.196123.250.90270.09135.814
5 DynEdge5.0186.744208.390.88870.094213.682
4 GRIT + DynEdge-D5.01833.348119.720.89940.067529.121
Fixed mixed specialists*3.4786.744116.570.89940.0675103.008

The mixed portfolio uses DeepIce (direction), DynEdge (vertex), ParticleNet (energy), and GRIT (topology and visible-y). It is not the cell-wise published-best envelope.

A visualization of these five-output trade-offs is provided in Figure 4, Appendix G (Discussion-stage external specialist comparison) of the revised supplementary material.

The results demonstrate a concrete accuracy–efficiency trade-off. PRISM achieves lower five-output serial pipeline latency than all three portfolios, while specialists deliver higher accuracy on several individual objectives. In particular, the mixed portfolio improves direction and vertex reconstruction, whereas PRISM achieves slightly higher topology AUC. These comparisons establish the computational benefits of PRISM under the evaluated serial inference setting, alongside the task-specific accuracy differences.

Low-energy reconstruction remains challenging. Beyond computational efficiency, PRISM investigates a central challenge in joint scientific reconstruction: how to share event-level information while preserving the task-specific evidence required by heterogeneous physical targets. Our task-subset experiments reveal non-monotonic transfer between objectives, while hit/skeleton retrieval ablations demonstrate their differing information requirements. These findings support PRISM's three-level design: shared event representation, task-aware information retrieval, and targeted structural specialization, and provide insights into information sharing and task differentiation in unified reconstruction.

Reviewer_3JFGOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PRISM a novel reconstruction framework that encodes sparse neutrino observations and encodes them into a single representation and decodes them into five representations for five different tasks: Topology classification, Neutrino Energy Estimation, Incoming Direction Estimation, Vertex localisation, Visible Inelasticity Classification. The proposed architecture only outperforms alternatives in 2 out of 5 tasks, on the main benchmark. The baselines were published at a NuBench benchmark with extensive numerical experiments and ablation studies

Overall Recommendation to AC: The paper is interesting highly technical on contemporary and actively developed topic that could be of interest to ML. However, the main claim of the value of "unification" is not really validated, so the core ML message is not supported by the results.

Critical strengths

Interesting and novel topic. The comparisons appear to be against state-of-the-art approaches and some of them were published in journals close to the ML community (e.g. GRIT published in ICML 2023). The method seems appropriate to the task.

Critical weaknesses

The main claim appears to be that having a common encoder will help in all five different tasks. To support that claim a main ablation study should report how the architecture performs at each of the tasks individually and then joint training to combine them.

Table 1 suggests that the common encoder helps in Energy and Topology but less so in Direction and Vertex. It is likely that the selection of tasks is competitively. Ablations with different subsets of taks might reveal interesting synergies

Table 1c seems silly. The authors execute the same neural network used for one task and all five tasks at once and claim it is five times more efficient, the result is 5 times slower by construction. Training five different neural networks for the different tasks with an aggregate parametrisation of the same size as PRISM would be the fair and appropriate comparison.

Questions for authors

Did you run Prism for each task individually? Did common train improve all results or only some?

Have you tried training on subsets of tasks, e.g. only Energy+Topology, and only Direction+Vertex?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3JFG

1. Single-Task versus Joint Training(Q1/W1):

We trained five independent single-task (ST) controls and a five-task joint PRISM model under matched Flower-XL conditions (seed 42, three epochs, effective batch 96), evaluating all models on the complete official test set (2,980,023 events). These supplementary checkpoints are separate from the submitted benchmark models.

Table A: Single-task vs. joint reconstruction (Flower-XL, full test set). Vertex errors are reported before and after matched Anchor refinement.

Training SettingTopoAUC↑Ewidth(%)↓Dir(°)↓VertexPre(m)↓VertexPost(m)↓yMedian↓
Separate-ST0.813689.54712.74815.34118.5450.1955
Five-task-joint0.905561.8327.76619.08213.1500.0891

Joint training provides task-dependent gains: energy width and direction error decrease by 30.95% and 39.08%, respectively, while initial vertex error increases by 24.39%. After matched Anchor refinement, joint PRISM achieves lower vertex error than the refined ST control. These results distinguish the initial task trade-off from the post-refinement outcome.

Paired event-bootstrap 95% CIs for Joint − ST are [−27.854, −27.572] pp for energy width, [−4.998, −4.967]° for direction, [+3.726, +3.759] m for initial vertex, and [−5.410, −5.380] m for refined vertex. These reflect event-sampling uncertainty for fixed checkpoints, not training-seed variability.

2. Task-Subset Interactions (Q2/W2):

We additionally trained Energy+Topology and Direction+Vertex subsets under matched Flower-XL conditions and evaluated validation-selected checkpoints on the complete official test set. Vertex results include the matched 20k-update Anchor refinement.

Table B: Task-subset comparisons on Flower-XL. ST and joint controls are identical to Table A. A dash indicates an unsupervised task.

Training SettingTopoAUC↑Ewidth(%)↓Dir(°)↓VertexPre(m)↓VertexPost(m)↓
Separate-ST0.813689.54712.74815.34118.545
Energy+Topology0.8287106.771———
Direction+Vertex——8.11617.55013.787
Five-task-joint0.905561.8327.76619.08213.150

The results reveal non-monotonic task interactions. Energy+Topology improves topology but increases energy width relative to ST, while Direction+Vertex improves direction but exhibits a different vertex trade-off. These results show that transfer depends on task composition, motivating task-aware information retrieval and specialization.

3. Independently Trained Specialists and Inference Efficiency (W3):

We agree that the original five-copy comparison primarily measures duplicated computation. We trained five independent smaller specialists and pruned unused branches, resulting in 6.117M deployed parameters, compared with 7.845M for joint PRISM.

Table C: Independently trained small specialists versus joint PRISM on Flower-XL. Vertex results are shown before/after matched 20k-update Anchor refinement. Latency measures sequential five-specialist versus joint five-output inference using post-refinement, pruned models. Best values are in bold.

SettingDeployedParams(M)TopoAUC↑Ewidth(%)↓Dir(°)↓VertexPre/Post(m)↓yMedian↓Latency(ms/event)↓
Five-small-ST6.1170.801190.95914.48019.058/22.5570.183833.497
Five-task-joint7.8450.905561.8327.76619.082/13.1500.08918.420

Under the evaluated five-specialist comparison, joint PRISM improves topology, energy, direction, and visible-y, with nearly identical initial vertex errors. After matched Anchor refinement, joint PRISM achieves lower vertex error, although the specialist deteriorates during refinement. The comparison demonstrates benefits of shared reconstruction under the evaluated protocol, rather than matched-capacity superiority.

Inference protocol. Latency was measured on an RTX A6000 (PyTorch 2.5.1/CUDA 12.4, FP32), using 128 fixed validation events, 50 warm-up calls, five repeated traversals, and at least 200 synchronized forward calls, excluding data loading and transfers. This protocol is separate from submitted Table 1(c).

Joint inference achieves a 3.98× FP32 speedup over sequential five-specialist execution at Batch 32. This demonstrates the computational benefit of shared five-task reconstruction under the evaluated deployment protocol.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.