{"id":"0a4712eb-d4ef-4b23-bbf5-9fdd61a3b9b0","arxiv_id":"2506.04622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A CNN trained on synthetic line shapes from a uniformized S-matrix classifies the CLAS Sigma-pi spectrum as the two-pole Lambda(1405) structure on the second Riemann sheet.","lead":"Researchers used a convolutional neural network to classify the shape of the Lambda(1405) peak in CLAS photoproduction data and concluded it is a two-pole structure on the second Riemann sheet. The work is an early demonstration of machine-learning-based line-shape analysis for an old hadron puzzle, though the method has acknowledged biases.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central inference rests on an untested representability assumption: the CLAS spectrum is assumed to lie in the 19-class synthetic family generated by F(sqrt(s)), so the CNN's 99% Label 04 confidence is a closed-set statement, not evidence for the physical two-pole structure.","rationale":"The reader's weakest assumption is exactly the representability of the CLAS line shape by the two-channel S-matrix generator in Section 2.2. My read of the paper does not change that assessment. The synthetic validation is competent: the precision, recall, and f1-scores on generated data are high and the confusion matrices are reported clearly. The paper also honestly notes that 19 pole structures are not exhaustive and that randomized label groupings should be explored. However, none of the reported metrics address whether the experimental spectrum is in-distribution for the trained CNN. A softmax classifier over a closed label set always outputs a label; 99% confidence only means the input resembles Label 04 more than the other 18 generated shapes. Since Stages 4 and 5 show Labels 13 and 14 are nearly degenerate with Label 04, small unmodeled effects such as background or resolution could plausibly flip the inference. Therefore the physical conclusion remains conditional. I would keep the verdict at CONDITIONAL and request the representability check plus randomized curriculum ordering as the concrete path to strengthen the claim.","tokens_in":5464,"tokens_out":8344,"duration_ms":114165,"concrete_test":"Augment the Section 2.2 generator with an explicit smooth non-resonant background term and a Gaussian detector-resolution convolution, retrain the CNN on the augmented label set (original 19 plus a background-only class), and re-run the CLAS inference on 10,000 resampled spectra. If the inferred label leaves Label 04, or if a binned likelihood-ratio test prefers the augmented model over the original Label 04 generator by Delta chi^2 > 4 per added parameter, the representability assumption is violated and the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All training and validation examples are generated in Section 2.2 from F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2, a two-channel uniformized S-matrix with pole terms only. The CLAS Sigma-pi spectrum is then assigned one of the 19 labels. The paper never checks whether the experimental data actually falls inside the distribution of any generated class: no goodness-of-fit, no energy-test, and no rejection option is provided. Because a softmax classifier is closed-set, it will pick a label even when the true spectrum contains non-resonant backgrounds, production-vertex effects, three-body phase space, or detector resolution not present in the generator. This is not a formal quibble: in Stages 4 and 5 the confusion matrices show Labels 13 and 14 (three-pole structures) are nearly degenerate with Label 04 at the 93-96% f1 level, so the line-shape differences among competing hypotheses are comparable to the distortions that background or resolution would introduce. The conclusion that Lambda(1405) is 'indeed' two-pole thus depends on an unvalidated representability assumption rather than on the CNN alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a convolutional neural network to classify line shapes of the Lambda(1405) in the Sigma-minus-pi-plus invariant mass spectrum into one of 19 pole-structure classes. The training and validation data are generated from a two-channel uniformized S-matrix with independent poles, using the line-shape generator F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2. After a six-stage curriculum training, the authors feed 10,000 randomized variants of the CLAS Sigma-minus-pi-plus data into the network and obtain label 04 (two poles on the second Riemann sheet) with 95--100% confidence in each stage. They conclude that Lambda(1405) is indeed a two-pole structure, consistent with the present consensus.","tokens_in":5753,"tokens_out":5770,"duration_ms":66379,"significance":"An ML classifier that can distinguish pole structures from line-shape data would be a useful tool for hadron spectroscopy, and the paper's synthetic-data validation is a genuine strength: most labels achieve precision, recall, and F1 scores above 0.9, and the curriculum strategy is a resource-efficient way to handle 19 classes. If the inference on the CLAS data were robust, it would support the two-pole picture with an independent line-shape analysis. However, the experimental inference is only as strong as the assumption that the CLAS spectrum lies within the 19-class generator family, and the paper does not test that assumption. The conclusion is therefore a model-based classification rather than an independent derivation.","major_comments":[{"comment":"The central inference depends on the assumption that the CLAS Sigma-minus-pi-plus invariant mass spectrum is representable by F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2 with the 19 pole structures of Table 1, but no goodness-of-fit or out-of-distribution test is reported. A softmax classifier is closed-set: it must return one of the 19 labels even when the input contains backgrounds, production-vertex effects, three-body phase space, or detector resolution that the generator excludes. The claim in Section 4 that Lambda(1405) is 'indeed' a two-pole structure is therefore not justified by the CNN output alone; please add a rejection class or an explicit comparison such as an energy distance or likelihood ratio between the CLAS binned distribution and the generated Label 04 samples.","section":"Section 2.2 and Section 2.4"},{"comment":"In the curriculum training, Label 04 is an anchor in every stage: Stage 1 outputs Label 04 from labels 01--04, and each subsequent stage re-trains with Label 04 plus three new labels. If Stage 1 makes a wrong inference, the error is propagated and later stages cannot recover because Label 04 is always present and the other labels are only ever compared with it. The paper acknowledges randomized label groupings only as future work, but without a sensitivity study the 95--100% confidence reported in Table 3 does not address this bias. Please rerun the curriculum with different anchor choices or without an anchor, and report how often the final inference changes.","section":"Section 2.3, Table 2"},{"comment":"Stages 4 and 5 show near-degenerate classification between Label 04 and Labels 13 and 14, with F1 scores of 0.94--0.95 and off-diagonal confusion in the confusion matrices. The line-shape differences among these hypotheses are thus comparable to the distortions that a small background or resolution effect would introduce, so the statement that the confusion is 'so minimal that it does not significantly affect the findings' is unsupported. Please report the full distribution of prediction counts for the 10,000 randomized CLAS inputs rather than only the maximum, and quantify how far the experimental inputs sit from the Label 04 versus Label 13/14 decision boundary.","section":"Section 3, Table 3"},{"comment":"The generator F(sqrt(s)) depends on r, theta, the pole positions omega_m, and the regulators omega_m', but the paper does not state the ranges, priors, or sampling distributions for these parameters. Without this information the training set is not reproducible, and one cannot tell whether the physical line shape could fall outside the support of the generated classes. Please specify the parameter ranges and show, if possible, that the CLAS data are covered by the generated Label 04 distribution.","section":"Section 2.2"}],"minor_comments":[{"comment":"The Figure 1 caption says 'energy and intensity as inputs' but the figure appears to show the CLAS invariant mass distribution, not the network architecture or input format; please clarify the figure and describe how inputs are preprocessed and normalized.","section":"Section 2.1 and Figure 1"},{"comment":"Please describe how the inferred label from the previous stage is used in the next stage: is the next model trained from scratch or fine-tuned, and how are the 10,000 randomized CLAS inputs combined with the synthetic data in that retraining?","section":"Section 2.3"},{"comment":"Check the expression omega = (q1+q2)/sqrt(epsilon_2^2 - epsilon_1^2); the subscripts on the thresholds are not defined consistently with q1 and q2, and the square root may be a typo.","section":"Section 2.2"},{"comment":"There are several typographical and formatting errors, such as 'In Stages 2 and 3„' and 'pure lineshape analysis'; please proofread the text.","section":"General"},{"comment":"The hyperparameters of the CNN (kernel sizes, number of channels, optimizer, learning rate, batch size) and the ranges used to randomize the CLAS energy and intensity values are not reported; please include them for reproducibility.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a proceedings contribution with a clear ML pipeline and strong synthetic validation, but the headline conclusion exceeds what the analysis can support because of the closed-set, model-generated training paradigm. I recommend major revision requiring at least one representability check and a sensitivity analysis of the curriculum anchor. If the authors soften the conclusion to 'among the 19 considered pole structures, the classifier selects the two-pole sheet-II label for the CLAS spectrum,' the paper could become acceptable after those additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the HADRON proceedings by Chavez and Sombillo. Short version: they train a CNN to classify the CLAS Sigma-pi invariant mass line shape into one of 19 pole structures from a two-channel uniformized S-matrix, and the network picks Label 04 (two poles on sheet II), which they read as confirmation of the Lambda(1405) two-pole consensus.\n\nWhat is actually new: this is the first application of a CNN (rather than a fully connected DNN) to this reaction, and the modified curriculum training with six four-label models is a reasonable way to limit resources while keeping accuracy. The synthetic validation is decent—most f1 scores above 0.9, and the confusion matrices are honestly reported.\n\nThe soft spot is the experimental inference. The training data are generated from F(s) = |T11 + r e^{-i theta} T21|^2, with pole terms only. The classifier is closed-set: it must pick one of the 19 labels. The paper never checks whether the CLAS data actually fall inside the distribution of any generated class—no goodness-of-fit, no energy test, no out-of-distribution rejection. So the 99% confidence in Label 04 is a statement about the 19-class family, not about the physical Lambda(1405). That is not a minor quibble; stages 4 and 5 show Labels 13 and 14 (three-pole structures) with f1 ~0.94-0.95 against Label 04, so the line-shape differences among competing hypotheses are comparable to the distortions a background or resolution effect could introduce.\n\nAlso: Label 04 appears as an anchor in every curriculum stage, which could bias the inference; the authors acknowledge this and propose randomized groupings for future work. And the free parameters (r, theta, pole positions, regulators) are not specified in the proceedings text, so the training set is not reproducible from the paper alone.\n\nCredit where due: the authors are transparent that they only consider 19 pole structures and call the results preliminary in the abstract. The final sentence 'indeed a two-pole structure' is stronger than the evidence supports, but that is typical of a conference write-up.\n\nWho this is for: someone working on ML applications in hadron spectroscopy, or on line-shape analysis near thresholds. The paper is a useful example of both the promise and the pitfalls of using classifiers for physics inference.\n\nPeer review: I would send it to a referee rather than desk reject, because the synthetic validation is meaningful and the method could transfer to other threshold enhancements. But the referee should require a representability check and the generation parameters before the physics claim is accepted. As a proceedings, it is fine as a preliminary report; as a journal paper, it needs that additional work.","headline":"A competent CNN validation on synthetic line shapes that overreaches in its experimental inference, but the method deserves a careful referee.","tokens_in":6265,"tokens_out":3134,"would_cite":false,"duration_ms":35894,"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":"The paper claims that a convolutional neural network classifies the measured $\\Sigma^-\\pi^+$ line shape of $\\Lambda(1405)$ as a two-pole structure on the second Riemann sheet, and takes this as agreement with the current consensus.","keywords":["Lambda(1405)","line shape analysis","convolutional neural network","two-pole structure","Riemann sheet","uniformized S-matrix","invariant mass distribution"],"falsifier":"Generate test spectra from the same 19 pole classes but add a smooth background or a third channel; if the CNN's confidence in Label 04 drops below its reported 95--100\\% level, the inference on the experimental spectrum is not robust to model misspecification. Alternatively, refit the measured spectrum with a one-pole and a two-pole generator and compare fit quality; a comparable one-pole fit would undercut the two-pole conclusion.","tokens_in":5269,"feed_emoji":"⚛️","tokens_out":9519,"duration_ms":80294,"temperature":0.7,"pith_summary":"The paper asks whether the $\\Lambda(1405)$ enhancement, whose nature has been debated since its prediction in 1960, really consists of two poles, as most recent analyses conclude. It builds a convolutional neural network that classifies line shapes by their pole structure, training it on 19 synthetic classes generated from a two-channel uniformized $S$-matrix that lets the pole positions and Riemann sheets be set by hand. When the measured $\\Sigma^-\\pi^+$ invariant mass distribution from the $\\gamma p\\to K^+\\Sigma\\pi$ reaction is fed through the network, the inference lands on the class with two poles both on the second Riemann sheet. The approach matters because peaks near a two-hadron threshold are ambiguous between kinematical effects and genuine dynamical states, and a classifier gives a way to test the pole interpretation without committing to a specific dynamical model.","feed_headline":"CNN reads Λ(1405) peak as two-pole structure","feed_subtitle":"Among 19 pole classes, the network pins the observed Sigma-pi peak to two second-sheet poles.","key_machinery":"The line-shape generator $F(\\sqrt{s}) = |T_{11}(\\sqrt{s}) + r\\,e^{-i\\theta} T_{21}(\\sqrt{s})|^2$ constructed from independent $S$-matrix poles is the object that carries the argument. The poles are placed through the uniformization variable $\\omega = (q_1+q_2)/\\sqrt{\\epsilon_2^2-\\epsilon_1^2}$, which gives control over the position and Riemann sheet of each pole and lets the authors generate 10,000 training line shapes for each of the 19 labels. The CNN itself has two convolutional layers followed by linear layers with ReLU activations; it is trained in six stages of four labels each, and its task is to map features such as the unitarity below the second threshold and the number and shape of peaks to the correct pole label.","core_discovery":"On the paper's own terms, the central discovery is that the experimental $\\Sigma^-\\pi^+$ spectrum selects Label 04 -- two independent poles on the second Riemann sheet -- among the 19 pole configurations in Table 1. The network reaches this answer with 95--100\\% confidence at each of the six curriculum-training stages, with the only notable confusions occurring between Label 04 and other two-pole labels in Stages 4 and 5. The paper presents this as agreement with the present consensus that $\\Lambda(1405)$ is a two-pole structure, one narrow pole near the $\\bar{K}N$ threshold and one broad pole near the $\\Sigma\\pi$ threshold.","pith_inferences":["A reader might infer that the network is effectively a non-parametric classifier for whether a spectrum shows a threshold-cusp pattern consistent with independent poles; its 'two-pole' verdict says nothing about the dynamical origin (molecular vs. quark) of those poles.","The same architecture could be pointed at other ambiguous states, such as the $P_{\\psi}^N(4312)^+$, to see whether a single training recipe separates kinematical cusps from genuine poles across different channels.","Because the training generator contains no background or three-body terms, the paper's result implicitly assumes those terms do not distort the measured spectrum; adding such terms to the generator is a direct test of whether Label 04 survives model misspecification."],"forward_implications":["If the classification is correct, the $\\Lambda(1405)$ is a two-pole structure on the second Riemann sheet, consistent with the current consensus of a narrow pole near the $\\bar{K}N$ threshold and a broad pole near the $\\Sigma\\pi$ threshold.","The trained CNN distinguishes all 19 pole classes with precision, recall, and F1-scores mostly above 0.9, showing that line-shape features carry enough information to separate one-, two-, and three-pole configurations.","The curriculum label groupings let the model extend to more complex pole structures without retraining earlier stages, so the same approach can be scaled beyond the 19 classes considered here.","The paper's program continues with the $\\Sigma^+\\pi^-$ and $\\Sigma^0\\pi^0$ invariant mass spectra, which would provide a cross-check of the two-pole inference in other final states."],"supporting_citations":[{"why":"introduces the independent S-matrix pole construction used to generate the 19 training line-shape datasets.","marker":"[10]"},{"why":"supplies the uniformization variable that controls pole positions and Riemann sheets in the line-shape generator.","marker":"[14]"},{"why":"provides the measured Sigma-minus-pi-plus invariant mass distribution that the CNN classifies.","marker":"[11]"},{"why":"pioneered deep-network classification of line shapes, establishing the method this CNN extends.","marker":"[7]"},{"why":"contributes the two-channel curriculum training scheme that the paper modifies into six stages.","marker":"[8]"},{"why":"one of the consensus two-pole analyses that the CNN inference is compared with.","marker":"[5]"},{"why":"another independent two-pole analysis supporting the second-Riemann-sheet interpretation.","marker":"[6]"}],"fun_headline_variants":["CNN sees Λ(1405) as two poles","CNN distinguishes two-pole Λ(1405) shape","Neural network classifies Λ(1405) as two poles","CNN backs two-pole Λ(1405)","CNN reads CLAS data as two Λ(1405) poles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured spectrum is assumed to be fully representable by the two-channel uniformized line-shape generator $F(\\sqrt{s})$ with independent poles; if backgrounds, three-body dynamics, or omitted channels shape the data, the CNN cannot detect them, and its choice of Label 04 would not be a valid inference about $\\Lambda(1405)$.","fun_headline_variants_meta":{"raw":{"variants":["CNN sees Λ(1405) as two poles","CNN distinguishes two-pole Λ(1405) shape","Neural network classifies Λ(1405) as two poles","CNN backs two-pole Λ(1405)","CNN reads CLAS data as two Λ(1405) poles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3213,"prompt_tokens":1013,"completion_tokens":2200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":2118}},"tokens_in":629,"tokens_out":2200,"duration_ms":16695,"temperature":1.0,"reasoning_tokens":2118,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:37:34.535947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate test spectra from the same 19 pole classes but add a smooth background or a third channel; if the CNN's confidence in Label 04 drops below its reported 95--100\\% level, the inference on the experimental spectrum is not robust to model misspecification. Alternatively, refit the measured spectrum with a one-pole and a two-pole generator and compare fit quality; a comparable one-pole fit would undercut the two-pole conclusion.","supporting_citations":[{"cited_title":"Interpretation of near-threshold peaks using the method of independent S-matrix poles","cited_arxiv_id":"2308.03325","evidence_quote":"introduces the independent S-matrix pole construction used to generate the 19 training line-shape datasets."},{"cited_title":"Kato,Ann","cited_arxiv_id":null,"evidence_quote":"supplies the uniformization variable that controls pole positions and Riemann sheets in the line-shape generator."},{"cited_title":"Spin and parity measurement of the Lambda(1405) baryon","cited_arxiv_id":"1402.2296","evidence_quote":"provides the measured Sigma-minus-pi-plus invariant mass distribution that the CNN classifies."},{"cited_title":"Sombillo, Y","cited_arxiv_id":null,"evidence_quote":"pioneered deep-network classification of line shapes, establishing the method this CNN extends."},{"cited_title":"Pole structure of $P_\\psi^N(4312)^+$ via machine learning and uniformized S-matrix","cited_arxiv_id":"2405.11906","evidence_quote":"contributes the two-channel curriculum training scheme that the paper modifies into six stages."},{"cited_title":"Meson-baryon reactions with strangeness -1 within a chiral framework","cited_arxiv_id":"1210.3485","evidence_quote":"another independent two-pole analysis supporting the second-Riemann-sheet interpretation."}],"review_version":1}