{"id":"5a5eb574-e0e5-4586-939a-aff371684533","arxiv_id":"2506.19920","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Sparse globally rigid graph representations of jets, combined with roughly 30 reclustered subjets, improve graph autoencoder anomaly detection on the LHC Olympics benchmark.","lead":"This paper builds a graph autoencoder that represents LHC jets as sparse rigid graphs and tests it on the LHC Olympics anomaly detection benchmark. It reports that an intermediate number of reclustered subjets and a specific sparse 'unique-6' graph give the best anomaly detection performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim rests on an unsupported comparison: no quantitative benchmark against prior autoencoder methods on LHCO is provided, and the reported max SIC is selected on the evaluation set.","rationale":"I read the rigidity concern raised by the reader as real but likely not fatal: the construction in Sec. 2.3 (start with K_{k+1}, repeatedly add a vertex with k edges to existing vertices, k>=3) preserves 3-connectivity and, generically, redundant rigidity, so Hendrickson's conditions are plausibly satisfied. The omitted proof is a clarity issue, not a correctness issue. The more weakly supported part of the central claim is the state-of-the-art assertion, which is a factual claim about the literature and about generalization, and the paper provides no evidence for either. The selection of the reported optimum on the evaluation set compounds this, making the headline max SIC an upper bound estimate rather than a validated performance. These gaps are directly testable and should be resolved before the SOTA claim is accepted.","tokens_in":23974,"tokens_out":18577,"duration_ms":196908,"concrete_test":"Compile the best-reported max SIC/AUC for unsupervised autoencoder-based methods on the LHC Olympics R&D dataset from the cited literature, and evaluate the fixed model (unique-6, nsubjets=30) on a held-out half of the official test sample that was not used in any hyperparameter scan. If any prior method reports max SIC >= 2.94 or AUC >= 0.925, or if the held-out max SIC is more than ~10% below the scanned maximum of ~3, the state-of-the-art claim and headline performance estimate require revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest claim (Sec. 3.3 and Conclusions: \"the performance of the unique-6 graph represents the current state-of-the-art among unsupervised autoencoder-based methods for this anomaly detection benchmark data set\") is not backed by any direct quantitative comparison. The authors cite autoencoder anomaly-detection works [19,20,24,25,26,27,28,30,31,55] but never tabulate their published max SIC or AUC on the same LHCO benchmark. The only comparison provided is to the authors' own variant with absolute node information, which reaches max SIC ~2, and to fully connected/Laman graphs. Thus a reader cannot verify whether unique-6 (max SIC ~3, AUC 0.925) actually exceeds all prior autoencoder methods. Additionally, the optimum configuration (nsubjets=30, unique-6, dlatent=2) was selected by scanning these hyperparameters on the same test sample used to report the final metric; the reported max SIC is the maximum over this scan and is upward-biased by selection. If a prior autoencoder already reaches max SIC >= 3, or if the selected-max drops under a proper holdout evaluation, the central headline claim is false or overstated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript develops a graph autoencoder for unsupervised anomaly detection at the LHC, using the LHC Olympics dataset as a benchmark. Jets are represented as graphs built from exclusive kT subjets, with node features given by transverse momenta and edge features by relative angular quantities. Several sparse connectivity structures are compared: locally rigid Laman graphs and globally rigid 'unique-k' graphs, alongside fully connected graphs. The authors report that the best performance occurs at an intermediate number of subjets (nsubjets ~25-30) with the unique-6 graph, giving a maximum SIC of approximately 3 and an AUC of 0.925, and they claim this is the current state-of-the-art among unsupervised autoencoder-based methods on this benchmark. The appendix extends the graph constructions to supervised jet classification tasks (quark vs. gluon, Z vs. QCD, top vs. QCD), finding that sparse graphs can approach the performance of fully connected graphs with far fewer edges. The code is publicly released.","tokens_in":24159,"tokens_out":13279,"duration_ms":132017,"significance":"If the main claims are substantiated, the paper makes a valuable conceptual contribution by linking graph rigidity theory to jet substructure and by demonstrating that sparse, globally rigid graphs can act as effective inductive biases for unsupervised anomaly detection. The empirical finding that performance peaks at an intermediate number of subjets and at an intermediate graph connectivity is interesting and well suited to the LHC Olympics benchmark. The release of the code and the careful description of the architecture are strengths that support reproducibility. The classification appendix is a useful additional result, showing that the graph constructions also illuminate supervised jet tagging. However, the significance is currently tempered by the absence of a direct quantitative comparison to prior autoencoder-based methods and by unresolved questions about the hyperparameter selection protocol for the headline numbers.","major_comments":[{"comment":"The statement that 'to the best of our knowledge, the performance of the unique-6 graph represents the current state-of-the-art among unsupervised autoencoder-based methods for this anomaly detection benchmark data set' is not supported by any quantitative comparison in the manuscript. The text cites Refs. [19,20,24,25,26,27,28,30,31,55] but never tabulates their published maximum SIC or AUC on the same LHC Olympics benchmark. The only comparisons shown are internal variants (absolute node information, fully connected graphs, modified Laman graphs). A reader cannot verify whether the reported max SIC ~3 actually exceeds all prior autoencoder methods. Please add a comparison table with published results from those references, ideally obtained with the same preprocessing and evaluation procedure, or soften the claim to a statement about the methods considered in this paper.","section":"Sec. 3.3 and Conclusions"},{"comment":"The manuscript does not state that the hyperparameters (nsubjets=30, unique-6, d_latent=2) were selected on a validation set rather than on the same 5e4 test events used to report the final metrics. Section 3.1 describes a validation set of 1e4 events and a test set of 5e4 events, while Section 3.2 says only that 'a hyperparameter scan showed that d_latent = 2 provides the best performance.' If the scan in Figs. 10 and 12 was evaluated on the test set, then the headline max SIC ~3 is the maximum over a scan of many configurations and is upward-biased by selection. Please clarify the exact protocol, and if the scan was performed on the test set, repeat the evaluation on a truly held-out set for the selected configuration, or apply a multiple-comparison correction and report the expected maximum under the null.","section":"Sec. 3.1-3.3, Figs. 10 and 12"},{"comment":"The latent representation is per-node with dimension d_latent, so the total latent size scales linearly with the number of subjets N. Varying nsubjets simultaneously changes the physical granularity of the input, the total number of input features (quadratically for fully connected graphs), and the total bottleneck width. The conclusion that anomaly-detection performance peaks at intermediate nsubjets ~25-30 may therefore reflect a capacity or input-dimensionality effect rather than the information content of subjet clustering. Please add a control experiment that keeps the total latent dimension approximately fixed, for example by scaling d_latent with 1/N or by using a graph-level bottleneck, or discuss this confound explicitly.","section":"Sec. 3.2, Eq. (3.2)-(3.3), Fig. 10"},{"comment":"The paper asserts that the proposed unique-k construction 'yields a graph that satisfies the Hendrickson conditions for global rigidity,' but no proof is given, and the cited Ref. [70] characterizes globally rigid graphs in general and does not establish that this particular nearest-neighbor addition sequence produces globally rigid graphs. Since the physical interpretation of the performance gain relies on global rigidity, please provide a proof or a precise citation showing that the construction is generically globally rigid for any sequence of nearest-neighbor choices, and ideally verify the 3-connectivity and redundant-rigidity conditions on the actual jet and subjet graphs used in the numerical study, which may contain near-degenerate angular separations.","section":"Sec. 2.3"}],"minor_comments":[{"comment":"The caption states a 'max SIC of 0.294' while the main text and the vertical axis of the figure indicate a value around 3; this inconsistency appears to be a typo and should be corrected.","section":"Fig. 10 caption"},{"comment":"The caption says 'Here the unique-30 graph is fully connected,' but for nsubjets=30 the fully connected graph corresponds to unique-29 under the definition in Sec. 2.3, since a unique-k graph starts from a clique on k+1 nodes. Please correct this inconsistency.","section":"Fig. 12 caption"},{"comment":"The manuscript uses S/B = 3% but does not specify how the training, validation, and test samples are constructed from the LHC Olympics data to achieve this signal fraction; please state the exact subsampling procedure.","section":"Sec. 3.1"},{"comment":"The claim that 'comparable performance is observed across a range of values with S/B <= 3%' is not accompanied by a figure or table; please provide the supporting data or remove the claim.","section":"Sec. 3.3"},{"comment":"The sentence 'See Refs. for more details [64, 99-127]' appears to have a missing citation marker; please fix the wording.","section":"Appendix A.1"},{"comment":"There are duplicate references in the bibliography: Ref. [83] duplicates [80], Ref. [38] duplicates [81], Ref. [84] duplicates [39], and Ref. [53] duplicates [31]. Please consolidate them.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline major revision rather than a rejection: the empirical study is careful, the code is released, and the central numerical findings may well hold. The two biggest risks are the unsupported state-of-the-art claim and the lack of a clear holdout protocol for the hyperparameter scan. If the authors can either provide a direct quantitative comparison to prior autoencoder results on LHCO or appropriately qualify the claim, and can clarify or correct the selection procedure, the paper would be suitable for publication. I would not require the rigidity proof to be fully formal, but an explicit statement of the generic assumptions and a numerical check on the actual jet graphs would substantially strengthen the conceptual argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThis is a solid paper about a real idea: using sparse graphs with rigidity constraints as the input representation for jet autoencoders. The central empirical finding—performance peaks for unique-6 graphs at n_subjets ~ 30 and beats fully connected graphs—is believable and the authors provide code and a clear description of the architecture. The appendix on jet classification is a useful control, showing that the same sparse graphs do not help supervised classification, which strengthens the anomaly-detection story.\n\nWhat is new here is the application of Laman and unique-k graph constructions to jet substructure, together with exclusive-kT subjet interpolation as a way to smoothly dial information content. The modified-Laman comparison in Fig. 12 is a good attempt to separate sparsity from topology, and the conclusion that the specific connectivity matters more than edge count is plausible.\n\nThe soft spots are real but not fatal. First, the state-of-the-art claim in Sec. 3.3 and the Conclusions ('to the best of our knowledge...') is not backed by any quantitative comparison to prior autoencoder methods on the LHC Olympics benchmark. Citing [19-32] is not the same as tabulating their max SIC or AUC. If a prior method already reaches max SIC >= 3, the headline is wrong. Second, the hyperparameters (n_subjets, k, d_latent) are scanned on the same test sample used to report the final metrics, so the reported max SIC is a selected maximum with upward bias. A separate validation split or nested procedure would fix this. Third, there is an internal inconsistency: the Fig. 10 caption says 'max SIC of 0.294' while the text and Fig. 12 say max SIC ~ 3. That looks like a factor-of-ten unit mistake, but it needs fixing.\n\nFourth, the rigidity guarantee is asserted rather than verified. The k-nearest-neighbor construction with pT ordering may not produce generically rigid unique graphs for real jet configurations with collinear particles; the Hendrickson conditions are sufficient for generic positions, and actual jets are not generic. The empirical success of unique-6 is encouraging, but the graph-theoretic interpretation is somewhat loose. Finally, only one signal topology is tested; the authors acknowledge this in the outlook.\n\nOverall, I think the paper is worth a serious referee. The central idea is novel, the work is reproducible, and the flaws are addressable. I would encourage the authors to add a direct comparison table, move the hyperparameter selection to a validation set, and soften or properly support the state-of-the-art claim. If that is done, this could be a nice JHEP paper.","headline":"Solid paper on sparse rigid graphs for jet anomaly detection; central finding is believable, but the state-of-the-art claim and test-set selection need work before I'd trust the headline number.","tokens_in":24731,"tokens_out":3355,"would_cite":false,"duration_ms":33711,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph autoencoder using 'unique-6' sparse graphs reaches max SIC ≈ 3 on the LHC Olympics benchmark, the best the authors know among unsupervised autoencoders.","keywords":["anomaly detection","graph autoencoder","graph rigidity","unique graphs","Laman graphs","LHC Olympics","jet substructure","unsupervised learning"],"falsifier":"On a sample of $10^4$ jets from the LHC Olympics benchmark with $n_{\\text{subjets}} = 30$, construct the unique-6 graphs and test each one for redundant rigidity and 3-vertex-connectivity; if a non-negligible fraction fail either condition, the graphs are not globally rigid in the claimed sense.","tokens_in":23697,"feed_emoji":"⚛️","tokens_out":12794,"duration_ms":118610,"temperature":0.7,"pith_summary":"This paper proposes that the connectivity of a graph representation of a jet can be used as a physics-informed knob for unsupervised new-physics searches, and demonstrates it with a graph autoencoder on the LHC Olympics benchmark. The authors construct sparse graphs with guaranteed geometric rigidity — locally rigid Laman graphs and globally rigid 'unique-k' graphs — and show that a unique-6 graph built from roughly 25 to 30 exclusive kT subjets gives the best anomaly-detection performance, with maximum SIC around 3 and AUC 0.925. They argue that this represents, to their knowledge, the current best result among unsupervised autoencoder-based methods on this data set. The same sparse graphs applied to supervised jet-classification tasks lose only a fraction of a percent of AUC relative to a fully connected model, so the benefit of rigid sparse graphs is specific to the unsupervised setting.","feed_headline":"Unique-6 graph hits SIC ~3 for LHC anomaly detection","feed_subtitle":"Sparse but globally rigid graphs, not just sparsity, are what let the autoencoder spot the new-physics signal.","key_machinery":"The key object is the unique-k graph construction. Nodes are particles or subjets ordered by decreasing transverse momentum; the $k+1$ hardest nodes form a fully connected clique, and every later node is joined to its $k$ nearest angular neighbours in the $\\eta$--$\\phi$ plane. For $k \\ge 3$ the result has $|E| = kN - k(k+1)/2$ edges, scales linearly with the number of nodes, and is globally rigid in $\\mathbb{R}^2$ because it satisfies Hendrickson's conditions of redundant rigidity and 3-connectivity. The autoencoder consumes node features $p_{T,i}$ and the edge features $\\theta_{ij} = (\\Delta\\eta_{ij}^2 + \\Delta\\phi_{ij}^2)^{1/2}$, $k_{T,ij} = \\min(p_{T,i}, p_{T,j}) \\theta_{ij}$, and $z_{ij} = \\min(p_{T,i}, p_{T,j})/(p_{T,i}+p_{T,j})$, compresses through a two-dimensional latent node representation, and reconstructs both node and edge features; the per-event sum of the two reconstruction errors is the anomaly score.","core_discovery":"At its core, the paper claims that a globally rigid sparse graph, whose edge lengths fix the positions of all nodes up to overall rotations and translations, is a better input representation for an unsupervised anomaly detector than either a fully connected graph or a merely locally rigid one. With jet constituents ordered by $p_T$, each new node in a unique-$k$ graph is attached to its $k$ nearest neighbours in the $\\eta$--$\\phi$ plane; for $k \\ge 3$ the result is globally rigid in $\\mathbb{R}^2$ and still has only $O(N)$ edges. Tested on the LHC Olympics dijet data with $S/B = 3\\%$, the graph autoencoder using a unique-6 graph and $n_{\\text{subjets}} = 30$ reaches max SIC $\\approx 3$ and AUC $= 0.925$, and the paper states that this is, to the best of its authors' knowledge, state-of-the-art among unsupervised autoencoder methods on this benchmark. A control with 'modified Laman' graphs, which add the same number of edges but without the rigidity guarantee, does not reproduce the peak, which the paper takes as evidence that the connectivity structure itself, not merely the number of edges, is responsible.","pith_inferences":["If the rigidity mechanism is generic, the unique-k construction should transfer to other anomaly-detection benchmarks and signal topologies; a test across signal variants with different intermediate masses would show whether the $n_{\\text{subjets}} \\approx 30$ optimum moves with the signal's angular scale.","Because the construction breaks permutation invariance by ordering on $p_T$, an extension with random or group-averaged tie-breaking could isolate whether the $p_T$ ordering itself, rather than the rigidity, carries part of the benefit.","A sharper falsifiable test would feed the autoencoder randomly rewired graphs with the same degree sequence as unique-6; if the SIC peak survives rewiring, the benefit comes from the degree distribution rather than from global rigidity."],"forward_implications":["At intermediate subjet counts, near $n_{\\text{subjets}} \\approx 25$--$30$, unsupervised graph autoencoders beat both hadron-level and heavily clustered inputs on the LHC Olympics benchmark; the optimal input is not the most detailed one.","The unique-6 connectivity pattern outperforms fully connected graphs of the same subjets, so adding more pairwise distances beyond global rigidity can reduce anomaly-detection sensitivity rather than help it.","Global rigidity, not sparsity by itself, is the relevant inductive bias: modified Laman graphs with comparable edge counts do not show the same peak, implying that graph topology should be considered when designing jet representations.","The same sparse constructions degrade supervised jet classification by only about 0.3--0.4% AUC relative to a fully connected attention-based classifier while using roughly an order of magnitude fewer pairwise distances."],"supporting_citations":[{"why":"Supplies the LHC Olympics 2020 benchmark data set, including the Z'→XY dijet signal and the background samples used for all anomaly-detection numbers.","marker":"[9, 10]"},{"why":"Defines the exclusive kT clustering algorithm used to recluster jets into a variable number of subjets, the knob that interpolates between high- and low-level inputs.","marker":"[59]"},{"why":"Introduces Laman graphs and minimal local rigidity in the plane, the starting point for the sparse graph constructions.","marker":"[69]"},{"why":"States Hendrickson's conditions for unique graph realizations, which the paper invokes to guarantee global rigidity of unique-k graphs.","marker":"[70]"},{"why":"Introduces the EdgeConv message-passing layer that forms the encoder and decoder blocks of the graph autoencoder.","marker":"[90]"},{"why":"Defines the Significance Improvement Characteristic curve used as the primary anomaly-detection performance metric.","marker":"[92]"},{"why":"Establishes that autoencoder-based anomaly detection is stable up to percent-level signal fractions, justifying the S/B = 3% setup.","marker":"[46]"}],"fun_headline_variants":["Graph rigidity beats sparsity for LHC anomaly spotting","Sparse rigid graphs sharpen LHC new-physics search","Graph autoencoder hits SIC 3 with unique-6 net","Globally rigid graphs, not sparsity, drive LHC sensitivity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the claim that connecting each new particle or subjet to its k nearest angular neighbours really does pin down the whole jet geometry uniquely from the edge distances; if realistic jet configurations admitted several different shapes with the same edge lengths, the rigidity argument would not be what produces the performance gain.","fun_headline_variants_meta":{"raw":{"variants":["Graph rigidity beats sparsity for LHC anomaly spotting","Sparse rigid graphs sharpen LHC new-physics search","Graph autoencoder hits SIC 3 with unique-6 net","Globally rigid graphs, not sparsity, drive LHC sensitivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2229,"prompt_tokens":994,"completion_tokens":1235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1164}},"tokens_in":610,"tokens_out":1235,"duration_ms":10957,"temperature":1.0,"reasoning_tokens":1164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:22:51.921180+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a sample of $10^4$ jets from the LHC Olympics benchmark with $n_{\\text{subjets}} = 30$, construct the unique-6 graphs and test each one for redundant rigidity and 3-vertex-connectivity; if a non-negligible fraction fail either condition, the graphs are not globally rigid in the claimed sense.","supporting_citations":[{"cited_title":"Multivariate discrimination and the Higgs + W/Z search","cited_arxiv_id":"1010.3698","evidence_quote":"Defines the Significance Improvement Characteristic curve used as the primary anomaly-detection performance metric."}],"review_version":2}