{"id":"6589ba53-a472-44b2-92c9-514a9ef829e1","arxiv_id":"2608.02735","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Lorentz-equivariant transformers outperform standard transformers for jet and flavor tagging whenever geometric 4-vector features dominate, and a 48M pretrained equivariant model matches far larger foundation models on top tagging.","lead":"This paper compares standard and Lorentz-equivariant transformer networks for tagging jets and heavy-flavor particles at the LHC, looking at both accuracy and computing cost. It finds that symmetry-aware networks win when particle directions matter, while plain transformers stay competitive when scalar features dominate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparison hinges on a shared scaling prescription whose per-architecture optimality is asserted, not shown.","rationale":"The reader correctly identified the shared scaling prescription and hyperparameter plateau as the weakest assumption. My concern is slightly more pointed: the paper's fairness argument depends not just on the plateau being 'approximately optimal' but on the plateau being equally flat and equally wide for all architectures. The text asserts this similarity but presents no tuning curves, so the one sentence carrying the fairness of the entire comparison is unverified. This is not an internal inconsistency, and the architecture implementations are genuinely optimized, so I do not recommend rejection. A CONDITIONAL verdict is appropriate: the central qualitative claim (L-GATr-slim wins at fixed cost when 4-vectors matter) is likely robust, but the quantitative Pareto-frontier rankings and the 'above a modest threshold' phrasing should be treated as provisional until the plateau claim is documented. I agree with the reader's choice of weakest assumption and with the CONDITIONAL verdict, though I would place slightly more weight on the per-architecture optimality of the schedule than on the absence of seed-level error bars.","tokens_in":25712,"tokens_out":2037,"duration_ms":20306,"concrete_test":"Report, for each architecture and dataset, a small grid of learning rates and weight decays (e.g., 3 settings each, centered on the Table 1 values) for s=0 and s=2 on ATLASTop, and show that the best-performing point on each plateau lies within the reported per-metric uncertainty of the Table 1 point. If a non-Table-1 setting improves any architecture by more than the bootstrap uncertainty band of the fitted curve, re-run the fixed-cost Pareto-frontier analysis with per-architecture tuned settings and report whether L-GATr-slim remains on the frontier.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is a performance ordering across architectures at fixed cost, yet the only bridge between 'raw network size' performance curves and 'fixed cost actually paid by a collaboration' is the common continuous scaling parameter s in Table 1. The paper states that the chosen learning rate and weight decay 'lie on an approximately optimal plateau that is similar for all tasks investigated,' but this assertion is not backed by tuning curves or seed-level repeats. If, for example, ParT at s=2 were trained with its own optimal learning rate and reached the performance of L-GATr-slim, or if the baseline transformer at s=2 were undertrained by a too-low learning rate, the ordering in Figures 3, 4, 7, and 8 could partly reflect tuning effort rather than the intrinsic value of Lorentz equivariance. In addition, the multiplicative form of the scaling prescription couples depth, width, and heads to s, so a single s changes multiple architectural axes at once; without a per-architecture optimality check, a suboptimal point on one axis (e.g., over-parameterized depth for the transformer) can masquerade as an architectural limitation. The conditional claim of the abstract is well-supported only if the common plateau is genuinely flat and equally wide for every architecture. The 'approximately optimal plateau' claim is the least secured load-bearing element.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a systematic scaling comparison of five transformer-based jet taggers—a baseline transformer, ParT, L-GATr, L-GATr-slim, and LLoCa-Transformer—on three public LHC datasets: ATLASTop (boosted top tagging), JetClass (multi-class boosted-object tagging), and JetSet (flavor tagging). The authors introduce a single continuous scaling parameter s that controls width, depth, and attention heads, train networks with s = -2,...,2, and fit scaling curves of the form T∞ + B/N^β to loss, AUC, and background rejection. They additionally optimize all implementations for inference cost (FLOPs, CPU/GPU time, memory, energy) and compare architectures at fixed cost. The central finding is that Lorentz-equivariant architectures outperform the baseline transformer whenever 4-momentum information is discriminative, with L-GATr-slim generally on the Pareto frontier, while all architectures perform similarly on the scalar-dominated JetSet flavor-tagging task. A final section pre-trains a 48M-parameter L-GATr-slim on JetClass and fine-tunes it on the top-tagging benchmark, reaching performance comparable to much larger OmniLearned models at lower training energy.","tokens_in":25913,"tokens_out":8486,"duration_ms":75351,"significance":"If the reported hierarchy is robust, this is the first controlled scaling study of Lorentz-equivariant transformers for both large-R jet tagging and flavor tagging, and it provides concrete, actionable guidance for LHC collaborations: adopt L-GATr-slim above a modest inference-cost threshold for boosted-object tagging, and retain standard transformers for very low latency budgets and scalar-heavy flavor tagging. The paper's strengths include the use of public external datasets with held-out metrics, a public code release, consistent application of efficiency improvements (torch.compile, mixed precision where applicable, sparse jet representations) to all compared architectures, and an honest discussion of the conditional nature of the equivariance benefit. The main open questions concern the statistical and tuning rigor of the comparison rather than its conceptual design.","major_comments":[{"comment":"The central comparison assumes that the shared hyperparameter prescription — in particular the channel-dependent learning rate and the common weight decay of 1e-2 — places every architecture on an approximately optimal plateau. The text states this in one sentence ('We validate these settings and find that they lie on an approximately optimal plateau') but provides no tuning curves, no per-architecture learning-rate scan, and no seed-level repeats. Because the scaling parameter s couples depth, width, heads, and (for ParT) the number of class-attention blocks, a single suboptimal hyperparameter choice for one architecture can shift the entire fixed-cost ranking. Please add per-architecture sensitivity checks, at minimum a learning-rate scan at two sizes (e.g., s=0 and s=2) on one dataset, and show that the main ordering in Figures 3, 4, 7, and 8 is stable under these variations.","section":"Sec. 2.1 (Table 1) and Sec. 3"},{"comment":"All performance points in the scaling curves appear to be single training runs per architecture and size; the bootstrap bands in Table 3 and the figures quantify only the fit uncertainty from resampling the same five points, not run-to-run stochasticity. Differences that drive the conclusions — L-GATr-slim versus L-GATr on ATLASTop, ParT catching up at large s on JetClass, the CPU-time crossover around 1 ms in Figure 8 — could be within seed noise. Please provide seed-level repeats (at least three, ideally five) for the architectures on the Pareto frontier and for the crossover points, and report the resulting error bars on the loss/AUC/rejection values.","section":"Sec. 3, Figures 3-9"},{"comment":"The asymptotic loss L∞ is an extrapolation of Eq. (3) from only five parameter points per architecture, and the reported uncertainties are obtained by bootstrapping those same five fitted points. The ranking of asymptotic losses (e.g., L-GATr-slim L∞ = 0.2196 vs ParT L∞ = 0.2222 on ATLASTop) is therefore sensitive to the assumed functional form, which is not tested. Please either restrict the hierarchy claims to the observed parameter range or add a sensitivity analysis (e.g., alternative fitting forms or dropping one point) to demonstrate robustness of the L∞ ordering.","section":"Sec. 3, Table 3"},{"comment":"The resilience study is summarized as showing that symmetry-aware networks improve performance without increasing systematic uncertainties, but the text itself reports a 'sharp drop in modeling uncertainties at s=2' for the LLoCa-Transformer and the standard transformer and states that this 'will require further study.' As written, this unexplained non-monotonicity undermines the claimed resilience advantage for the largest networks. The authors should either resolve the drop with additional analysis or qualify the resilience conclusion to the sizes where it is supported.","section":"Appendix B, Figures 15-17"},{"comment":"The variance-preserving rescaling replaces the frame transformation L by L/γ with a per-particle γ_i defined in Eq. (11). Since γ_i varies across tokens, this is not a global rescaling: it changes the relative attention logits in Eq. (10) compared to the original LLoCa attention, and therefore constitutes a new attention mechanism rather than a pure stabilization. Figure 12 shows training stability but not final performance or attention behavior. Please compare the original and rescaled LLoCa on at least one benchmark (e.g., JetClass at s=0 and s=2) to demonstrate that the modification does not alter the model's expressivity or its ranking in the comparison.","section":"Appendix A, Eq. (10)"}],"minor_comments":[{"comment":"There are several typos, e.g., 'datatsets' and 'netwroks' in Appendix B and 'LLoCa-Tranformer' in Section 4; please correct them.","section":"Throughout"},{"comment":"Section 3.3 excludes ParT from the JetSet comparison because its implementation does not handle the large number of scalar features. This is understandable, but since ParT is one of the two main non-equivariant baselines, the claim that 'standard transformers' match equivariant performance on flavor tagging rests on the baseline transformer alone; the text should state this caveat explicitly when drawing the JetSet conclusion.","section":"Sec. 3.3"},{"comment":"Table 4 reports 'Uncertainties in AUC are at most ±0.0001 and not reported'; for consistency and to allow readers to judge the new s=2 entry, please report the AUC uncertainties or mark all AUC values as having negligible uncertainty.","section":"Table 4"},{"comment":"In Section 2.2, the sentence 'For L-GATr the amount of FLOPs is reduced by 60M for L-GATr dense because of the sparse geometric product' is confusing; Table 2 suggests the 60M reduction applies to L-GATr_dense and the 1650M reduction to L-GATr_sparse. Please clarify which implementation each number refers to.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own implementations of L-GATr and LLoCa, and while the code is public and the efficiency improvements appear carefully done, an independent reimplementation or a check by the original authors of ParT would strengthen the fairness of the cost comparison. Several references are to very recent arXiv preprints that may not yet be peer-reviewed (e.g., Refs. [12], [20], [21], [22], [24]); the authors should ensure the most stable versions are cited. The unexplained s=2 modeling-uncertainty drop in Appendix B is the kind of finding that a referee should ask to be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is the most comprehensive cost-aware benchmark of Lorentz-equivariant transformers for jet tagging I've seen. It compares five architectures across three public datasets (ATLASTop, JetClass, JetSet) at fixed inference cost, and the central claim—equivariant transformers win when 4-vectors dominate, and match baselines when scalar features matter—is well supported by the data. The paper also ships concrete improvements: optimized L-GATr sparse/dense implementations, a variance-preserving LLoCa attention rescaling, and a 48M-parameter L-GATr-slim that after pretraining on JetClass matches OmniLearned-M/L on top tagging. Code is public. That's real, reproducible work.\n\nWhat's genuinely new: the unified three-dataset scaling comparison, the cost-metric analysis (FLOPs, CPU/GPU time, memory, energy), and the pretraining result. The authors are careful to distinguish raw performance from performance at fixed budget, and the Pareto-frontier plots are useful.\n\nThe soft spots, in order of importance. The common scaling prescription (continuous parameter s in Table 1 plus shared LR schedule and weight decay 1e-2) is asserted to be on an 'approximately optimal plateau' but no tuning curves or seed-level repeats are shown. If the plateau isn't equally flat for all architectures, some of the ordering could reflect tuning effort rather than equivariance per se. That's a legitimate concern, and it's the weakest link in the quantitative ranking. Second, individual scaling points appear to be single training runs; the bootstrap on the fit gives uncertainty on the asymptote but not on the points. Third, ParT is excluded from JetSet because the implementation doesn't handle the 19 scalar features, so the flavor-tagging comparison lacks the strongest non-equivariant transformer. Fourth, energy use is modeled from operation counts, not measured, and they say real energy can be 100x larger. None of these overturn the conditional claim—the qualitative hierarchy is consistent across loss, AUC, and background rejection, and the JetSet result (all architectures similar) is exactly what you'd expect when scalars dominate. But they limit how much weight you can put on close rankings between, say, L-GATr-slim and LLoCa.\n\nWho it's for: anyone choosing a tagger architecture for an LHC analysis, and people working on physics-informed ML. It's worth a serious referee—send it out, but ask the authors to document the LR/weight-decay validation (or provide seed repeats for at least a few points) and to report per-point uncertainties. I'd review it myself if asked.","headline":"A solid, cost-aware benchmark showing Lorentz-equivariant transformers win when 4-vectors matter, with a few documentation gaps that should be fixed before the quantitative rankings are trusted.","tokens_in":26476,"tokens_out":1896,"would_cite":true,"duration_ms":18009,"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":"Lorentz-equivariant transformers beat standard taggers whenever 4-vector geometry matters.","keywords":["Lorentz equivariance","jet tagging","flavor tagging","transformer","scaling laws","inference cost","L-GATr","LHC machine learning"],"falsifier":"Train every architecture with an individual hyperparameter search on the JetClass dataset and compare test loss at fixed measured GPU inference time; if the baseline transformer reaches parity with L-GATr-slim after per-architecture tuning, the paper's ranking is an artifact of the shared recipe. Alternatively, rerun the ATLAS top-tagging benchmark with all models implemented through the same custom CUDA kernels; if the baseline's curve moves onto the Pareto front, the equivariant advantage was mostly implementation headroom.","tokens_in":25452,"feed_emoji":"⚛️","tokens_out":10281,"duration_ms":108928,"temperature":0.7,"pith_summary":"This paper tries to establish a conditional rule for LHC jet tagging: building the known Lorentz symmetry directly into a transformer pays off exactly when the discriminating information lives in particle four-momenta, and stops paying off when scalar summary features dominate. The authors compare five transformer families on three public datasets, scale each by a common recipe from 40 thousand to 40 million parameters, and track FLOPs, CPU and GPU time, memory, and energy use. Their finding is that the slim Lorentz-equivariant transformer leads the fixed-cost Pareto front for boosted-object tagging, while the ordinary transformer remains best for very short latency budgets and ties for flavor tagging, where impact-parameter significances are the decisive scalars. A reader should care because this is a direct architecture prescription for experiments preparing trigger and offline taggers under tight inference budgets.","feed_headline":"Equivariant transformers beat standard taggers on 4-vector jets","feed_subtitle":"Across three datasets and five cost metrics, Lorentz equivariance pays off when momentum vectors carry the signal.","key_machinery":"The comparison is built on four devices: a continuous scaling parameter $s$ that grows channels, blocks, heads, and equivariant feature multiplicities in lockstep across all architectures; a set of inference-cost metrics measured on CPUs and GPUs; a power-law fit $T(N)=T_\\infty + B/N^\\beta$ that separates asymptotic performance from convergence speed; and the geometric representations themselves. L-GATr operates on the 16-dimensional multivector representation of the Lorentz group, L-GATr-slim on a reduced vector-only subalgebra, and LLoCa on local Lorentz frames built by a small equivariant network plus Gram-Schmidt orthonormalization; each equivariant model also carries learned symmetry breaking, since the LHC beam and detector break the full Lorentz group. A new variance-preserving rescaling of the frame transformations, replacing $L$ by $L/\\gamma$, keeps deep LLoCa transformers stable at large width. These mechanisms jointly turn the question of whether equivariance helps into a set of cost-performance curves.","core_discovery":"The paper's stated conclusion is that Lorentz-equivariant architectures win whenever 4-vectors are relevant, and match the baseline performance when scalar information is important. In the ATLAS top-tagging study and the JetClass multi-class study, the equivariant L-GATr and its slim variant produce lower loss and higher background rejection than the baseline transformer and the approximately equivariant Particle Transformer, and the ordering survives once loss is plotted against inference FLOPs, energy, time, and memory rather than parameter count. In the JetSet flavor-tagging study, where track impact-parameter significances are the dominant scalar features, all architectures converge to nearly the same loss, and the baseline transformer sits on the Pareto frontier at low cost. Finally, a 48-million-parameter L-GATr-slim pretrained on 100 million JetClass jets and fine-tuned on top tagging matches the tagging performance of much larger billion-jet foundation models while using far less training energy.","pith_inferences":["One implication the authors leave open is a hybrid design: feed four-momentum streams through a Lorentz-equivariant trunk and scalar track features through a conventional trunk, then merge late; the JetClass versus JetSet contrast predicts this should outperform either pure design when both feature types matter.","A testable extension from the flavor-tagging null result is that equivariance will only pay off there if the task forces the network to combine kinematics, for example by adding an auxiliary vertex-mass or displaced-decay-length reconstruction objective.","The scaling-exponent differences suggest that increasing the training dataset may widen the equivariant advantage on JetClass while the baseline saturates, but this extrapolates beyond the data the paper fits.","Because the LLoCa stabilization term becomes important only at large depth, the paper predicts that shallow comparisons of equivariant and non-equivariant transformers will systematically understate the equivariant benefit at scale."],"forward_implications":["On boosted-object tagging, a collaboration facing a FLOP, energy, or memory budget should expect the slim Lorentz-equivariant transformer to dominate the standard transformer at every budget above a very modest floor.","A CPU latency budget below roughly one millisecond is the one regime where the baseline transformer is the safer choice for large-radius tagging.","For flavor tagging, adding Lorentz equivariance is not a free performance win; the deciding input is the set of scalar track significances, not the four-momenta.","Pre-training an equivariant transformer on a 100-million-jet sample can deliver the tagging performance of a foundation model trained on a billion jets, at a fraction of the training energy.","On the ATLAS top-jet dataset, the equivariant models' better rejection comes with at most a few percent increase in modeling uncertainty and no clear increase in experimental uncertainty, so the gain is not an artifact of simulation overfitting."],"supporting_citations":[{"why":"Supplies the L-GATr architecture, a Lorentz-equivariant transformer on the geometric-algebra multivector representation, which is the main equivariant model in the comparison.","marker":"[19]"},{"why":"Introduces the slim vector-only L-GATr variant that leads the fixed-cost comparisons, together with the energy-cost estimation used for the energy metric.","marker":"[20]"},{"why":"Introduces the LLoCa local-frame equivariant transformer and the baseline transformer implementation that this paper improves and scales.","marker":"[21]"},{"why":"Defines the Particle Transformer and provides the JetClass multi-class dataset used for large-radius tagging and for pre-training.","marker":"[11]"},{"why":"Provides the JetSet flavor-tagging dataset and the Salt-style transformer baseline that defines the scalar-dominated regime.","marker":"[35]"},{"why":"Supplies the ATLAS top-tagging dataset with full detector simulation and the systematic-variation samples used for the resilience study.","marker":"[39]"},{"why":"Motivates the power-law scaling fit that assigns each architecture an asymptotic loss and a convergence exponent.","marker":"[12]"},{"why":"Provides the billion-jet foundation-model baselines whose tagging performance the pretrained equivariant model is measured against.","marker":"[22]"},{"why":"Frames the comparison as learning a broken symmetry versus learning the symmetry breaking, the conceptual question the paper quantifies.","marker":"[25]"}],"fun_headline_variants":["Equivariant transformers win only when 4-vectors matter","Lorentz-equivariant taggers beat baselines on jet data","Equivariance pays off only when momentum vectors signal","Lorentz nets win on 4-vector jets, tie on scalar features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the single scaling prescription and shared hyperparameters place every architecture on an equally good operating point; if the common learning rates, weight decay, and parameter-scaling rule happen to favor the equivariant models, the measured hierarchy would reflect tuning luck rather than the value of Lorentz equivariance.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant transformers win only when 4-vectors matter","Lorentz-equivariant taggers beat baselines on jet data","Equivariance pays off only when momentum vectors signal","Lorentz nets win on 4-vector jets, tie on scalar features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000129,"raw_usage":{"total_tokens":1049,"prompt_tokens":801,"completion_tokens":248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":174}},"tokens_in":417,"tokens_out":248,"duration_ms":3862,"temperature":1.0,"reasoning_tokens":174,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:19.839650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train every architecture with an individual hyperparameter search on the JetClass dataset and compare test loss at fixed measured GPU inference time; if the baseline transformer reaches parity with L-GATr-slim after per-architecture tuning, the paper's ranking is an artifact of the shared recipe. Alternatively, rerun the ATLAS top-tagging benchmark with all models implemented through the same custom CUDA kernels; if the baseline's curve moves onto the Pareto front, the equivariant advantage was mostly implementation headroom.","supporting_citations":[],"review_version":1}