{"id":"1ab634af-8ef6-4ea0-92ad-82e8783ce3b3","arxiv_id":"2505.08736","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single transformer with separate pixel and time vocabularies generates realistic Cherenkov detector hits and supports particle identification and noise filtering after fine-tuning.","lead":"This paper builds a transformer model that simulates the pixel positions and arrival times of Cherenkov photons a detector would record for pion and kaon tracks, conditioned on momentum and angle. It also shows the same model can be fine-tuned for particle identification and noise filtering, with code released on GitHub.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported generation fidelity is sensitive to sampling parameters selected on the target distributions; at 6 GeV/c the KDE separation is below Geant4 and NF, so the 'high-fidelity' claim is not yet established.","rationale":"The reader's conditional verdict is appropriate, but the weakest assumption named by the reader, namely ordering ambiguity, is not the most load-bearing issue. Sorting hits by time gives a canonical order, so next-token prediction over time-sorted sequences is well-posed in principle; ties from time binning are a second-order nuisance. The paper's own statement about 'highly ambiguous' next-token prediction is made in the context of sampling sensitivity rather than ordering. The more consequential weakness is that the central quantitative evidence for 'high-fidelity generation' is a KDE separation at 6 GeV/c that falls below both the Normalizing Flow baseline and Geant4 (Fig. 6), and the generations are produced with T=1.05 and p=0.995 selected to match the target distributions. This makes the headline claim fragile: without a sensitivity analysis, one cannot tell whether the model learned a coherent joint distribution or whether the reported agreement is a product of tuned sampling. I would therefore keep the conditional verdict, with conditions that the authors report a sampling-parameter sweep and soften the 'high-fidelity' wording in the abstract until the 6 GeV/c closure is at least competitive with the NF baseline. The proposed concrete test directly settles this.","tokens_in":19826,"tokens_out":6779,"duration_ms":66863,"concrete_test":"Using the released checkpoints, regenerate the 6 GeV/c pion and kaon samples with a sweep of decoding parameters, e.g., T in {0.8, 1.0, 1.05, 1.2} and p in {0.9, 0.95, 0.995, 1.0}, and recompute the FastDIRC KDE separation of Fig. 6 for each setting. If the 6 GeV/c separation varies by more than ~0.3 s.d. across the sweep, or if the standard T=1.0/p=1.0 setting falls below the reported 2.57-2.59, the reported fidelity is sampling-dependent and the 'high-fidelity' claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of fast, high-fidelity generation rests on the cumulative distribution ratios (Fig. 4) and the FastDIRC KDE separation (Fig. 6). The quantitative closure at the central momentum is weaker than both references: at |p|=6 GeV/c the transformer variants reach 2.55-2.59 s.d., versus 2.91 for the Normalizing Flow baseline and 3.72 for Geant4 (Fig. 6, bottom panel). The paper acknowledges this: 'our proposed method does degrade in performance in comparison to both Geant4, and previously proposed methods such as NF' (Section 4, KDE based Evaluation). The favorable cumulative distributions are obtained with nucleus sampling p=0.995 and temperature T=1.05, which the authors state 'have shown to provide the most consistent cumulative distributions across the phase-space' (Section 4, Generative Model Evaluation), and they note that 'small changes in these parameters can result in significantly different outputs.' Because the sampling hyperparameters appear to be selected on the target distributions used for evaluation, and no sensitivity analysis or statistical uncertainty is reported for the distribution plots, the observed agreement could be a consequence of tuning rather than of a learned joint distribution. The ordering-ambiguity concern raised by the reader is related but secondary: sorting hits by time makes the autoregressive target well-defined, and time-bin ties are a second-order nuisance; the immediate gap is that the measured fidelity at the central momentum is below the published baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an autoregressive transformer for generating low-level Cherenkov detector readouts—pixel indices and arrival times—for the High Performance DIRC at the future EIC. The architecture introduces four claimed innovations: separate discrete vocabularies for space and time fused through Causal Multi-Head Cross-Attention (CMHCA), continuous kinematic conditioning via prepended context embeddings, fine-grained continuous-variable tokenization without joint vocabulary inflation, and class-conditional generation via a Mixture of Experts (MoE). The model is trained on Geant4-generated track-level hit sequences and evaluated through cumulative distribution ratio plots, photon-yield comparisons, and FastDIRC KDE separation power at fixed momenta, with additional downstream tasks of pion/kaon identification and noise filtering. The central claim is that a single (proto) foundation model can generate kinematically conditioned Cherenkov hit patterns with high fidelity and transfer to reconstruction tasks.","tokens_in":20096,"tokens_out":7143,"duration_ms":69079,"significance":"If the approach holds, it offers a useful step toward fast, conditional simulation of pixelated Cherenkov detectors without the resolution loss of VQ-VAE tokenization, while sharing a single backbone across particle classes and downstream tasks. The split-vocabulary CMHCA idea is interesting and plausibly transferable to other detector systems. The paper is honest in places: it publicly releases code, compares against Geant4 and a normalizing-flow baseline, and explicitly acknowledges some performance degradation. However, the main quantitative evidence for generation fidelity is mixed: at 6 GeV/c the transformer variants achieve KDE separation of 2.55–2.59 s.d., below both the NF baseline (2.91 s.d.) and Geant4 (3.72 s.d.), and the sampling hyperparameters appear to be selected on the target distributions without a sensitivity analysis. The significance is therefore moderate and contingent on addressing the open generation-fidelity and robustness questions.","major_comments":[{"comment":"At |p|=6 GeV/c the transformer variants reach 2.55–2.59 s.d. separation, which is below both the Normalizing Flow baseline (2.91 s.d.) and Geant4 (3.72 s.d.). This contradicts the abstract's 'high-fidelity generation' characterization. Please either soften the central claim, provide a quantitative argument for why this gap is acceptable (e.g., systematic uncertainty or downstream physics impact), or add additional metrics that support the high-fidelity claim. The text acknowledges degradation, but the abstract and conclusions do not reflect it, and this discrepancy is central to the paper's main claim.","section":"Section 4, KDE based Evaluation (Fig. 6, bottom)"},{"comment":"The sampling parameters T=1.05 and p=0.995 are explicitly selected because they 'have shown to provide the most consistent cumulative distributions across the phase-space', and the paper states that small changes can lead to significantly different outputs. Because the cumulative distribution ratios are the primary quantitative closure evidence, the reported agreement may be a result of tuning rather than a learned joint distribution. Please provide a sensitivity scan over T and p, report how the cumulative ratios and the KDE separation power vary, and specify a validation protocol that avoids selecting sampling parameters on the evaluation distributions.","section":"Section 4, Generative Model Evaluation"},{"comment":"The paper states that 'the idea of next token prediction of Cherenkov photons is highly ambiguous', yet the model is trained as an autoregressive next-token predictor over time-sorted hit sequences. Since the physical hit set is permutation-invariant, the factorization imposes a conditional structure that may not be unique or well-posed. Please justify why the learned conditional distribution over the sorted sequence correctly represents the joint hit distribution, and provide evidence (e.g., permutation-invariance or set-level generation metrics) that the generation quality is robust to the ordering convention and to time-bin tie-breaking.","section":"Section 4, Generative Model Evaluation (next-token ambiguity)"},{"comment":"The four claimed innovations are not isolated by ablation. For example, no comparison is shown against a single-joint-vocabulary transformer, a VQ-VAE tokenization baseline, or a model without CMHCA under otherwise identical training conditions. Without these ablations, the paper does not establish which components are responsible for the reported behaviour. Please add ablations or explicitly temper the innovation claims to what the current comparisons support.","section":"Section 3, Methods (split vocabularies and CMHCA)"}],"minor_comments":[{"comment":"The cumulative distribution ratios in Fig. 4 (and Appendices A/B) are shown without statistical uncertainties; please add uncertainty bands or state that the deviations are within statistical noise.","section":"Section 4, Histogram Level Evaluations"},{"comment":"The momentum labels '3 Gev/c' and '6 Gev/c' contain a typo; they should read '3 GeV/c' and '6 GeV/c'.","section":"Section 4, Particle Identification / Fig. 8"},{"comment":"The statement that the model 'achieves the desired separation power (3σ at 6 GeV/c)' refers to the PID task (Fig. 8), but at generation time the KDE separation in Fig. 6 is below 3σ at 6 GeV/c; please clarify which result is being cited.","section":"Section 5, Summary and Conclusions"},{"comment":"The sentence 'we see our model is able to efficiently able to characterize detector noise' contains a duplicated 'able to'; please correct.","section":"Section 4, Noise Filtering"},{"comment":"The claim that the autoregressive model 'more effectively captures the kaleidoscopic effect' is based on a single visual comparison in Fig. 7; please either quantify this effect or present it as qualitative evidence.","section":"Section 4, Geometric Effects"},{"comment":"The abstract and introduction emphasize 'fast' generation, but no runtime or computational-cost comparison with Geant4 or NF is reported; please add a timing measurement or remove the speed claim.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies on the authors' own prior work [18] for the dataset, evaluation metric, and NF baseline, which is appropriate but should be acknowledged clearly in the paper. The 'foundation model' framing is somewhat ambitious given the single-detector scope, though the 'proto' qualifier helps. I see no ethical concerns. The main technical risk is the sampling-parameter selection issue in Major Comment 2; if the authors can show robustness across hyperparameters, the contribution is sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a real methods contribution, not a hype piece. The split-vocabulary idea is new to detector fast simulation: instead of VQ-VAE tokens that smear continuous values, they keep pixel indices discrete and bin time linearly at finer resolution, then fuse the two streams with causal cross-attention where time queries space. Prepended kinematic embeddings and class-conditional MoE routing let one shared transformer generate pions and kaons. They also show fine-tuning to PID works, and noise filtering works when trained from scratch, and they are candid that fine-tuning does not help for token-level filtering. Code is public.\n\nThe evaluation is mostly honest. They compare against Geant4, a normalizing flow, and their own independent models, and they report the negative result: at |p|=6 GeV/c the KDE separation for the transformer variants is 2.55–2.59 s.d., below NF's 2.91 and Geant4's 3.72. That gap is real and it is not hidden. The cumulative distributions look good in x/y and reasonable in time, though the tail beyond 100 ns is off—which they dismiss as negligible because densities are low. Photon yield as a function of polar angle matches well.\n\nThe soft spots are the usual ones. Sampling temperature T=1.05 and nucleus p=0.995 were selected to give 'the most consistent cumulative distributions across phase-space,' and they say small changes produce significantly different outputs, but we get no sensitivity analysis and no statistical uncertainty on the distribution plots. Part of the closure could be tuning rather than a learned joint distribution. Also, the claim that VQ-VAE tokenization causes problematic resolution loss is asserted, not demonstrated: there is no VQ-VAE baseline trained on the same data. That would be the natural ablation. The ordering ambiguity of autoregressive Cherenkov hits is acknowledged, and sorting by time makes the target well-defined, so I do not see that as fatal.\n\nMy overall read: the core architectural ideas are worth taking seriously, the paper is well written, and the limitations are not hidden. But the headline 'high-fidelity' should be softened given the 6 GeV/c result, and the benchmark against a VQ-VAE tokenizer should be added before the central claim is established. A serious referee should engage with this—conditional acceptance with required revisions.\n\nBest.","headline":"A serious methods paper with a genuinely new split-vocabulary attention scheme for detector readout; the architecture is worth knowing, but the 'high-fidelity' claim is overstated at the kinematic point where it matters most.","tokens_in":20670,"tokens_out":2265,"would_cite":true,"duration_ms":21717,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single transformer with split pixel/time vocabularies can generate Cherenkov hit patterns for a future EIC detector and transfer to pion/kaon identification and noise filtering.","keywords":["foundation models","Cherenkov detectors","next-token prediction","causal cross-attention","mixture of experts","particle identification","fast simulation","noise filtering"],"falsifier":"Regenerate the published closure-test kinematics with temperature T=1.0 and nucleus probability p=1.0, so the sampler draws directly from the trained distribution, and compare the joint pixel-time distribution to the reference simulation; a sharp degradation there would show that the reported fidelity depends on hand-tuned sampling rather than on a learned joint distribution.","tokens_in":19559,"feed_emoji":"⚛️","tokens_out":9326,"duration_ms":87922,"temperature":0.7,"pith_summary":"This paper argues that one autoregressive transformer, trained with separate token vocabularies for pixel location and arrival time, can act as a proto-foundation model for pixelated detector readouts at the future Electron-Ion Collider. The core proposal is to avoid lossy learned tokenization: keep native pixel indices as discrete tokens, bin continuous arrival times at a fraction of the sensor resolution, and fuse the two streams with causal multi-head cross-attention where each time token queries its plausible pixel locations. Kinematic conditions, momentum magnitude and polar angle, are prepended as continuous embeddings, and a class-conditioned mixture of experts lets a single model generate both pion and kaon patterns. Closure tests on the High Performance DIRC show near-unity spatial ratios, photon yields that track the reference simulation across polar angle, pion/kaon separation meeting the 3-sigma requirement at 6 GeV/c, and token-level noise filtering with roughly 90 percent average precision. If the approach holds, detector simulation and reconstruction for Cherenkov systems could share one trainable backbone.","feed_headline":"One transformer fast-simulates Cherenkov hits and IDs particles","feed_subtitle":"Split pixel/time vocabularies plus class-routed experts match reference hit patterns and reach the 3-sigma pion-kaon bar.","key_machinery":"The load-bearing mechanism is the split-vocabulary causal sequence with Causal Multi-Head Cross-Attention (CMHCA): time embeddings are the query stream and pixel embeddings the key/value stream, so each arrival time queries its plausible pixel locations while the mask preserves autoregressive causality. Two linear heads then predict the next token over each vocabulary, avoiding any joint space-time vocabulary of roughly 36 million tokens; the model instead learns 6,144 pixel tokens and 5,920 time tokens and combines them by attention. $\\ell^2$-normalized query and key matrices with a learned scale, plus pre-layer normalization, keep attention focused rather than collapsed onto the most populated configurations across phase space. For class-conditional generation, the transformer's feed-forward networks are replaced by a mixture of experts with fixed class routing, so pion and kaon sequences pass through different expert subsets inside one shared backbone.","core_discovery":"The central claim is that next-token prediction over a time-sorted sequence of detector hits is a viable generative target for Cherenkov readout at detector-level resolution, provided space and time live in separate vocabularies and are fused by causal cross-attention. The pixel vocabulary is the native detector readout, 6,144 indices; time is linearly binned at one-quarter of the sensor resolution, giving 5,920 tokens. The model predicts the next pixel and the next time jointly, with two cross-entropy losses, conditioned on prepended continuous embeddings of momentum magnitude and polar angle. Replacing the feed-forward layers with class-routed mixture-of-experts layers produces generations that the paper finds in strong agreement with independent per-class models, so pions and kaons can share one model. Closure metrics show spatial distributions in ratio near one and photon yields in good agreement, while separation power reaches the 3-sigma EIC requirement at 6 GeV/c but remains below the reference simulation and the normalizing-flow baseline; the paper attributes the gap to inherent smoothing and to the freedom in choosing generation parameters.","pith_inferences":["A stronger test than the published marginals would compare full joint pixel-time correlations between generated and reference tracks; divergence there would indicate the success rests on tuned sampling, not on a faithful learned joint distribution.","An ablation that randomizes hit order within tracks during training would directly probe whether the autoregressive factorization is learning the hit set or the dataset's specific ordering; if it collapses, the ordering assumption, not the architecture, is carrying the result.","If the split-vocabulary pattern transfers, a detector stream could chain time, energy, and space through sequential cross-attention blocks, which the paper sketches as future work and which would extend the method to calorimeters without VQ-VAE resolution loss.","The low-momentum PID advantage suggests a momentum-weighted loss could reallocate capacity toward the high-momentum region where pion/kaon separation is hardest; the paper lists this as future work without testing it."],"forward_implications":["A single class-conditioned model can generate multiple particle species in one backbone, since closure tests show MoE generations agree with independent per-class models.","Because the spatial vocabulary is the native pixel readout, generated samples preserve per-pixel geometric structure, including kaleidoscopic patterns near the expansion volume that a continuous normalizing-flow generator smooths away.","The same pretrained backbone fine-tunes into a pion/kaon classifier with faster convergence and slightly higher accuracy than training from scratch.","Token-level noise filtering reaches about 0.90 average precision and about 0.99 rejection AUC, so the readout model can also tag dark-rate noise hits in high-rate EIC environments.","Directly learning sequence length removes the need for post-hoc photon-yield modeling as a function of phase space, a step previous fast Cherenkov simulation required."],"supporting_citations":[{"why":"Supplies the hpDIRC dataset of pion and kaon tracks, the normalizing-flow baseline, and the evaluation procedure the closure tests follow.","marker":"[18]"},{"why":"The prior next-token calorimeter simulation that motivates the split-vocabulary change by showing VQ-VAE resolution loss and missing conditional generation.","marker":"[13]"},{"why":"Defines VQ-VAE tokenization, the approach the paper avoids in order to preserve detector resolution.","marker":"[14]"},{"why":"Provides the cross-attention mask idea for combining separate sequences, used here as time-queries-space CMHCA.","marker":"[15,16]"},{"why":"Gives the transformer self-attention block that processes the fused spatio-temporal embeddings.","marker":"[20]"},{"why":"Prior Cherenkov reconstruction with the delta-log-likelihood method, the particle-identification baseline compared in Fig. 8.","marker":"[17]"},{"why":"Supplies the KDE-based separation metric used to measure generated-sample fidelity in the closure tests.","marker":"[28]"},{"why":"Nucleus sampling with fixed temperature is the generation procedure used in all reported samples.","marker":"[30]"},{"why":"Focal loss handles the class imbalance in the token-level noise-filtering task.","marker":"[25]"}],"fun_headline_variants":["Fast Cherenkov simulation and pion-kaon ID with one transformer","Cross-attention joins pixel and time tokens for Cherenkov fast-sim","Mixture-of-experts Cherenkov sim: one model for hits and PID","One model, two vocabularies: transformer Cherenkov fast-sim","Cherenkov transformer hits 3-sigma pion-kaon separation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That ordering a track's Cherenkov photons by time and predicting them one token at a time is a well-posed way to learn the joint hit distribution, even though the order of individual photons is not physically meaningful; the paper itself calls next-token prediction of Cherenkov photons 'highly ambiguous' and relies on hand-tuned sampling temperature and nucleus probability (T=1.05, p=0.995) to keep generations coherent.","fun_headline_variants_meta":{"raw":{"variants":["Fast Cherenkov simulation and pion-kaon ID with one transformer","Cross-attention joins pixel and time tokens for Cherenkov fast-sim","Mixture-of-experts Cherenkov sim: one model for hits and PID","One model, two vocabularies: transformer Cherenkov fast-sim","Cherenkov transformer hits 3-sigma pion-kaon separation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00116,"raw_usage":{"total_tokens":4809,"prompt_tokens":957,"completion_tokens":3852,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":3752}},"tokens_in":573,"tokens_out":3852,"duration_ms":26573,"temperature":1.0,"reasoning_tokens":3752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:48:01.292381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regenerate the published closure-test kinematics with temperature T=1.0 and nucleus probability p=1.0, so the sampler draws directly from the trained distribution, and compare the joint pixel-time distribution to the reference simulation; a sharp degradation there would show that the reported fidelity depends on hand-tuned sampling rather than on a learned joint distribution.","supporting_citations":[{"cited_title":"Neural discrete representation learning 2017 Advances in neural information processing systems 30","cited_arxiv_id":null,"evidence_quote":"Defines VQ-VAE tokenization, the approach the paper avoids in order to preserve detector resolution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the transformer self-attention block that processes the fused spatio-temporal embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior Cherenkov reconstruction with the delta-log-likelihood method, the particle-identification baseline compared in Fig. 8."},{"cited_title":"FastDIRC: a fast Monte Carlo and reconstruction algorithm for DIRC detectors","cited_arxiv_id":"1608.01180","evidence_quote":"Supplies the KDE-based separation metric used to measure generated-sample fidelity in the closure tests."}],"review_version":1}