Pith. sign in

REVIEW 4 major objections 6 minor 14 references

GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding an active-ingredient safety graph to medication recommendation improves accuracy and cuts the drug-drug interaction rate by about half.

desk verdict The ingredient-level DDI granularity is a genuinely new idea and the matched two-row comparison is well framed, but the paper's central attribution claim is unsupported until the missing ablation, the λ values, and the variability details are actually provided. read the letter →

arxiv 2608.00098 v1 pith:UOA7NGQJ submitted 2026-07-30 q-bio.QM cs.AI

Juao Fan , Jinhan Li , Shengxin Zhu This is my paper
classification q-bio.QMcs.AI
keywords medicationrecommendationdrug-druginteractionactiveingredientgranularitystatespacemodelelectronichealthrecordspolypharmacysafetymulti-labelpredictionnormalization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the active ingredient, not the packaged drug code or the molecular substructure, is the right granularity for enforcing drug-drug interaction safety in medication recommendation. It presents a recommender that encodes patient histories with a linear-time state space sequence model and trains with a joint objective over three graphs: a drug-level DDI graph, an ingredient-level DDI graph built by normalizing drug codes to active ingredients, and an EHR co-prescription graph. In a strictly matched comparison where the only difference is the ingredient-level signal, the model improves on all multi-label accuracy metrics while roughly halving the drug-level DDI rate. The authors conclude that ingredient-level normalization recovers predictive signal lost when codes aggregate multiple ingredients, and that safety modeling at this granularity is complementary to accurate sequence modeling.

What carries the argument

The load-bearing object is the ingredient-level DDI graph A_ING: medication codes are normalized to active ingredients via RxNorm, ingredient-pair interactions are taken from TwoSIDES, and the graph is projected onto the medication vocabulary for use as a quadratic penalty on predicted probabilities. The same ingredient graph, used natively (Algorithm 2), defines the ingredient-level DDI evaluation metric. A proportional controller adjusts the accuracy-safety trade-off to the observed validation DDI rate.

What would settle it

Compute the ingredient-level DDI rate for the baseline model using the same ingredient normalization and checking pipeline (Algorithm 2) and compare to IngMamba's 0.2063; if the baseline is already at or below that value, the claimed safety benefit of the ingredient-level signal disappears. Alternatively, train with a random held-out subset of ingredient-pair edges removed and test whether the model generalizes to those held-out pairs.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that activating the ingredient-level DDI graph and its associated loss terms—while leaving the backbone, cohort, preprocessing, and evaluation code untouched—raises Jaccard from 0.4488 to 0.4983, PRAUC from 0.6911 to 0.7485, and F1 from 0.5989 to 0.6453, and lowers the drug-level DDI rate from 0.1875 to 0.0948 on MIMIC-IV. The paper reads this as evidence that the ingredient-level signal supplies information that the drug-code-level graph does not, and it introduces an ingredient-level DDI rate computed in native ingredient space as a safety metric that code-level evaluation cannot see.

Load-bearing premise

The ingredient-level DDI graph built from RxNorm and TwoSIDES is accurate enough to serve as both the training penalty and the evaluation metric; if that graph is biased or the drug-to-ingredient mapping is wrong, the reported improvement could be an artifact of the graph rather than genuine safety gain.

Editorial extensions

If this is right

  • If the claim holds, medication recommenders should standardize safety knowledge at the active-ingredient level rather than the drug-code level.
  • Ingredient-level DDI evaluation should be reported alongside drug-level rates, since code-level metrics miss interactions between ingredients hidden inside aggregated codes.
  • The result suggests the accuracy-safety trade-off is not intrinsic: aligning the safety penalty with the granularity of clinical knowledge can improve both at once.
  • The linear-time sequence backbone means longer, irregular ICU trajectories can be modelled without quadratic cost, which the paper shows does not conflict with the safety gains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A stricter test of the safety claim would compute the baseline's ingredient-level DDI rate under the same normalization pipeline; the paper leaves this unreported, so we do not yet know whether the ingredient penalty reduces ingredient-pair risk or merely shifts risk to pairs outside the graph.
  • Because the evaluation metric is computed on the same graph used for training, part of the measured 'improvement' could reflect overfitting to the knowledge base. Holding out a random subset of ingredient-pair interactions and testing generalization would settle this.
  • The paper's controller targets a 0.05 drug-level DDI rate but only reaches 0.0948; raising the safety weight could close the gap, and the expected accuracy cost is a testable prediction.
  • If ingredient-level normalization is the real driver, models trained on drug codes with denser DDI graphs might see similar gains; the paper's controlled comparison isolates the ingredient signal, but replicating the result with a different DDI knowledge base would test its generality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes IngMamba (also referred to as GRAIN in the title-level abstract), a medication recommendation model for MIMIC-IV built on a selective state-space backbone and a multi-granular safety objective. Medication tokens are standardized to ATC3, while an additional ingredient-level DDI graph is constructed via RxNorm and TwoSIDES and is used both as a training penalty and as an evaluation metric. Under a controlled two-row comparison against a re-implemented MambaHealth baseline, the paper reports consistent accuracy gains (Jaccard 0.4488 to 0.4983, PRAUC 0.6911 to 0.7485, F1 0.5989 to 0.6453) and a drop in drug-level DDI rate from 0.1875 to 0.0948. The central claim is that the ingredient-level signal is responsible for these gains because the two rows differ only in whether the ingredient-level graph and its loss terms are active.

Significance. If fully substantiated, the ingredient-level normalization idea is valuable: DDI knowledge is curated at ingredient granularity, and recovering signal lost in code-level aggregation could improve both accuracy and safety. The controlled matched-settings comparison is a sound starting point, and the two DDI evaluation algorithms are clearly specified. However, the paper currently does not substantiate its headline attribution or its safety conclusions: the key ablation is missing, the loss weights are TODO, the safety metrics reuse the same graphs that appear in the training objective, and the variability reporting is incomplete. The paper is candid about several of these limitations, which is commendable, but the central empirical claims are not yet supported by the presented evidence.

major comments (4)
  1. [§4.5.1, Eq. (18)] The paper's headline claim — that the improvement is attributable to the ingredient-level signal — is not established. The two-row comparison varies the joint activation of the λING and λEHR terms, not the ingredient-level component alone; the λ values are left as TODO; and §5 explicitly states that no component-wise ablation of the three graph terms is reported. The dynamic controller (Eq. 19) can also drive different β trajectories in the two runs, further confounding attribution. Please provide ablations with λING on/off while holding λEHR fixed, report the actual λ values, and report the resulting β trajectories.
  2. [§4.1.2–4.1.3 vs. §3.5.1] The safety evaluation is circular relative to the training objective. Eq. (23) computes the drug-level DDI rate using A_DDI from ddi_A_final.pkl, which is the same matrix that enters LDDI in Eq. (12); Eq. (24) checks predicted ingredient pairs against G, the same graph used to construct AING in Eq. (13). The reported DDI reductions therefore measure the model's ability to suppress the exact safety signal it was trained on, not an independent forecast of interaction risk. Even granting the paper's caveats about TwoSIDES bias, this same-graph evaluation cannot support the safety claim. Please add a held-out DDI evaluation on a knowledge source not used in training, and report Ingredient-DDI for the baseline as well.
  3. [§4.2, Table 2] The variability intervals are not informative: both models report identical spreads (±0.002/±0.003), and the method for computing intervals is listed as TODO. Without a stated number of seeds or bootstrap details, the reader cannot tell whether the reported gains of +0.0495 Jaccard and −0.0927 DDI rate are within run-to-run noise. Please report mean ± std over at least five seeds with the same protocol for both models.
  4. [§4.5.2, Table 3] The ingredient-level DDI rate is reported only for IngMamba; the baseline row is '—'. The paper therefore makes no comparative ingredient-level safety claim, and the text appropriately calls the number a reference value. However, the abstract and contribution list also claim 'a more clinically plausible average medication count per visit', but no such number appears anywhere; the TODO in §4.5.2 explicitly acknowledges this. Either add the required columns (Avg. #Med and baseline Ingredient-DDI) or remove/soften those claims.
minor comments (6)
  1. [Title/Abstract] The paper title and abstract use 'GRAIN' while the full text uses 'IngMamba'. Please unify the naming.
  2. [§3.1] The dataset statistics table is a TODO: no #patients, #visits, avg visits/patient, vocabulary sizes, or ingredient/DDI-edge counts are provided. These numbers are needed to assess cohort and graph coverage.
  3. [Figures] Figures 1–4 are placeholders ('file not found'); the training curves and architecture diagrams mentioned in the text are not available in the manuscript.
  4. [§4.2] The MIMIC-IV version is listed as TODO. Since preprocessing and cohort characteristics differ across versions, this must be specified.
  5. [§5 / Data Availability] The public code repository URL is TODO, so the paper's reproducibility claim is currently aspirational. Please release code or state that it will be provided in the final version.
  6. [§4.4] DATR is cited as an Anonymous ICLR 2026 submission and used for context. Please ensure the reference is updated with the final venue and authors, or remove it if anonymity cannot be preserved.

Circularity Check

2 steps flagged · score 6.0 of 10

DDI-rate 'improvement' is the training objective itself; accuracy anchor is external but ingredient-level attribution is admitted to lack ablation.

  1. self definitional [§3.3.3 and §3.5.1 Eq. (13) vs §4.1.3 Eq. (24)]
    "Two views of this graph are used, and it is important to distinguish them: • For training, we project G onto the medication vocabulary to obtain A_ING, where A_ING_ij = 1 if any ingredient of token i interacts with any ingredient of token j in G. This projection makes A_ING dimensionally compatible with the probability-based safety scores of Section 3.5. • For evaluation, we operate in the native ingredient space: a predicted medication set is decomposed into its ingredient set and pairs are checked directly against G (Algorithm 2)."

    The training loss r_ING_t = s·p_t^T A_ING p_t (Eq. 13) is built from the same graph G that defines the evaluation metric Ingredient-DDI (Eq. 24), which counts predicted pairs (g_i,g_j) ∈ G. The model is explicitly trained to minimize probability mass on G-edges, so the reported Ingredient-DDI = 0.2063 is the value of the training objective at threshold, not an independent safety forecast. Any bias in TwoSIDES/G is inherited by both the loss and the metric.

  2. self definitional [§3.5.1 Eq. (12) vs §4.1.2 Eq. (23)]
    "For the drug-level DDI adjacency, r_DDI_t = s·Σ_i Σ_j p_t,i p_t,j A_DDI_ij = s·p_t^T A_DDI p_t ... Following MambaHealth and prior work, the drug-level DDI rate uses the released DDI matrix ddi_A_final.pkl."

    The drug-level DDI training loss L_DDI is exactly the quadratic form p_t^T A_DDI p_t, and the evaluation metric Drug-DDI (Eq. 23, Algorithm 1) counts predicted pairs against the same adjacency A_DDI (ddi_A_final.pkl). Thus the headline reduction from 0.1875 to 0.0948 is the expected effect of optimizing the training objective, not an out-of-sample prediction. The paper calls this 'the central empirical claim' (§4.5.2), but it is by construction a report of the optimized quantity.

full rationale

The paper has one genuine circular element: the reported safety improvements are computed on the same knowledge graphs that appear in the training losses. Drug-level: Eq. (12) minimizes p_t^T A_DDI p_t, while Eq. (23)/Algorithm 1 score predicted sets against the same A_DDI (ddi_A_final.pkl). Ingredient-level: Eq. (13) minimizes p_t^T A_ING p_t where A_ING is the projection of G, while Eq. (24)/Algorithm 2 score predicted sets directly against G. So the DDI reductions are optimized quantities, not independent predictions, and biases in the external resources are inherited by both the loss and the metric. The accuracy gains (Jaccard 0.4488→0.4983, PRAUC 0.6911→0.7485, F1 0.5989→0.6453) against the re-implemented MambaHealth baseline are a genuine external anchor, since they are not directly optimized by the ingredient graph; this prevents a higher score. However, the paper itself flags that the central attribution is unsupported: §5 states 'we also do not currently report ... nor a component-wise ablation of the three graph terms; both are needed to fully substantiate the motivation for the SSM backbone and the attribution of gains to the ingredient-level signal,' and the TODO after Table 3 says 'Without it, reviewers cannot attribute the gain to the ingredient-level component specifically, which is the paper's headline contribution.' The λ values in Eq. (18) are also TODOs, so the exact objective is not fully specified. There is no load-bearing self-citation chain or imported uniqueness theorem; the MambaHealth citation is simply the baseline. Overall score 6: partial circularity, with the safety claim reducing to the training objective while the accuracy claim retains independent content.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

GRAIN/IngMamba's central claim rests on a long chain of hand-set choices: ATC3 tokenization, frequency filters, DDI graph filters, loss weights (some TODO), controller constants, and external knowledge bases (RxNorm, TwoSIDES, STITCH). The safety result is evaluated on the same graphs used to train; no ablation isolates the ingredient-level term; no code or parameters are released.

free parameters (8)
  • λDDI, λING, λEHR (loss weights) = TODO (unspecified)
    Eq. 18 defines the unified objective but leaves the knowledge-source weights as placeholders; central model behavior depends on them.
  • s scale factor = 5e-4
    Eq. 12/13; chosen to keep safety terms comparable to the task loss.
  • task loss blend α = 0.95
    Eq. 14; weights BCE against multi-label margin loss.
  • dynamic safety controller (γ, Kp, βmin, βmax, m) = 0.05, 0.5, 0.2, 0.95, 0.9
    Eq. 19–20; adapts β to validation DDI rate; all constants are hand-set.
  • medication tokenization and frequency filters = top 2000 Dx, 1000 Px, 300 ATC3 meds
    Section 3.1; vocabulary pruning affects which DDIs can be predicted.
  • DDI graph filters = top 60 side effects, freq≥15, ≤1000 CID pairs
    Section 3.2.2; post-hoc filtering of STITCH determines the A_DDI used for training and evaluation.
  • SSM state dimension d_state = |Vm|
    Section 3.4.2; chosen empirically, no sensitivity analysis.
  • decision threshold τ = 0.5
    Eq. 11; inference threshold, no tuning analysis.
assumptions (5)
  • domain assumption MIMIC-IV retrospective prescriptions are a valid proxy for clinical quality
    Evaluation treats agreement with recorded prescriptions as accuracy; the Limitations section acknowledges this proxy.
  • domain assumption TwoSIDES/RxNorm DDI knowledge is sufficiently complete and unbiased for training and evaluation
    G is built from these resources; the paper notes reporting biases and unknown DDIs, so low DDI rates only mean avoidance of known risks.
  • ad hoc to paper Dropping NDC codes that fail to map to ATC3 does not bias the cohort
    Section 3.1 drops unmapped medications; if unmapped codes carry risk signal, results are skewed.
  • ad hoc to paper ATC3 is the correct token granularity for preserving both coverage and safety signal
    Design choice with no sensitivity analysis; the central claim relies on it.
  • ad hoc to paper Filtered STITCH graph (top 60 side effects, freq≥15, 1000 pairs) captures the relevant DDI structure
    Post-hoc filters define the graph used in training and evaluation; changing them likely changes the reported rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation." pith.science (2026). https://pith.science/paper/UOA7NGQJ

@misc{pith2026260800098,
  author       = {Pith},
  title        = {Pith review of: GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOA7NGQJ}},
  note         = {Machine review of arXiv:2608.00098}
}
read the original abstract

Medication recommendation from electronic health records must balance predictive accuracy against the risk of adverse drug-drug interactions (DDIs) under polypharmacy. Existing safety-aware recommenders operate at one of two granularities: the drug code, which treats each medication as an indivisible token, or the molecular substructure, which is finer than pharmacological interaction knowledge is actually organized. We argue that the active ingredient is the missing granularity, and introduce GRAIN, a medication recommendation framework built around it. GRAIN encodes longitudinal patient trajectories (diagnoses, procedures, past medications) with a selective state space backbone that handles long, irregular visit sequences in linear time. On top of it we introduce a joint objective unifying three knowledge sources aligned to a common medication vocabulary: a drug-level DDI graph, an ingredient-level DDI graph obtained by normalizing medication codes to active ingredients via RxNorm, and an EHR-derived co-prescription graph. A proportional controller adapts the accuracy-safety trade-off to the observed validation DDI rate rather than fixing it a priori. Under strictly matched settings -- identical preprocessing, cohort, vocabulary, split, and evaluation code -- GRAIN improves over a re-implemented MambaHealth baseline on MIMIC-IV across all standard multi-label metrics (Jaccard 0.4488 to 0.4983, PRAUC 0.6911 to 0.7485, F1 0.5989 to 0.6453) while reducing the drug-level DDI rate from 0.1875 to 0.0948. We further define an ingredient-level DDI rate, a safety measure invisible to drug-code-level evaluation. The results indicate that ingredient-level normalization recovers predictive signal erased by code-level aggregation, and that it is complementary to, rather than in competition with, accurate sequence modeling.

Figures

Figures reproduced from arXiv: 2608.00098 by the authors.

Figure 1
Figure 1. MIMIC-IV preprocessing and medication token standardization pipeline (NDC [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Two graphs aligned to the medication vocabulary: (A) the EHR co-prescription graph [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. End-to-end framework and model architecture of IngMamba: a two-stream shared Mamba [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Training dynamics over epochs for IngMamba (ingredient-aware) and the MambaHealth baseline [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

  1. [1]

    DATR:DDI-Awaretherapeuticstructurereconstructionforsafermedicationrecommendation (ICLR 2026 submission)

    Anonymous. DATR:DDI-Awaretherapeuticstructurereconstructionforsafermedicationrecommendation (ICLR 2026 submission). OpenReview, September

    2026

  2. [5]

    16 Alistair E

    URLhttps://arxiv.org/abs/2111.00396. 16 Alistair E. W. Johnson, Tom J. Pollard, Lu Shen, Li-wei H. Lehman, Mengling Feng, Mohammad Ghassemi, others, and Roger G. Mark. MIMIC-III, a freely accessible critical care database.Scientific Data, 3:160035,

    arXiv

  3. [10]

    Stuart J

    URLhttps://arxiv.org/abs/2309.05675. Stuart J. Nelson, Kelly Zeng, John Kilbourne, Tamara Powell, and Robin Moore. Normalized names for clinical drugs: RxNorm at 6 years.Journal of the American Medical Informatics Association, 18(4):441–448,

    arXiv

  4. [14]

    URLhttps://doi.org/10.24963/ijcai.2021/514

    doi: 10.24963/ijcai.2021/514. URLhttps://doi.org/10.24963/ijcai.2021/514. 18

    work page doi:10.24963/ijcai.2021/514 2021

  5. [2011]

    URL https://doi.org/10.1136/ amiajnl-2011-000116

    doi: 10.1136/amiajnl-2011-000116. URL https://doi.org/10.1136/ amiajnl-2011-000116. Karan Singhal, Tulika Tu, Julius Gottweis, Rory Sayres, Ellery Wulczyn, Lily Hou, others, and Vijay Natarajan. Large language models encode clinical knowledge.Nature, 620:172–180,

    doi:10.1136/amiajnl-2011-000116 2011

  6. [2012]

    Mambahealth: Alightweightfoundationmodelforefficientdrug recommendation

    YudaWang,XuxinHe,andShengxinZhu. Mambahealth: Alightweightfoundationmodelforefficientdrug recommendation. InAIM-FM Workshop, 38th Conference on Neural Information Processing Systems (NeurIPS 2024),

    2024

  7. [2016]

    URLhttps://www.nature.com/articles/ sdata201635

    doi: 10.1038/sdata.2016.35. URLhttps://www.nature.com/articles/ sdata201635. Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Aya Shammout, Steven Horng, Tom J. Pollard, Sicheng Hao, Benjamin Moody, Benjamin Gow, Li-wei H. Lehman, Leo A. Celi, and Roger G. Mark. MIMIC-IV, a freely accessible electronic health record dataset.Scientific Data...

    doi:10.1038/sdata.2016.35 2016

  8. [2017]

    URLhttps://doi.org/10.1093/jamia/ocx010

    doi: 10.1093/jamia/ocx010. URLhttps://doi.org/10.1093/jamia/ocx010. Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

    arXiv

Show all 14 references
  1. [2018]

    URL https://doi.org/10.1145/3219819.3219981

    doi: 10.1145/3219819.3219981. URL https://doi.org/10.1145/3219819.3219981. Chuan Liu, Jia Lin, Jun Wang, Hong Liu, and James Caverlee. Mamba4rec: Towards efficient sequential recommendation with selective state space models. arXiv,

  2. [2020]

    URLhttps://journals.plos.org/plosone/article?id= 10.1371/journal.pone.0235353

    doi: 10.1371/journal.pone.0235353. URLhttps://journals.plos.org/plosone/article?id= 10.1371/journal.pone.0235353. KinWahFung,JeannineKapusnik-Uner,JoannaCunningham,SusanHigby-Baker,andOlivierBodenreider. Comparison of three commercial knowledge bases for detection of drug–drug...

    doi

  3. [2021]

    ThiagoNunesGonzagadeAndradeSantos,SilviaAlvesMartinsMendonça,MarceloGurgelCarlosdaSilva de Oliveira, and Helena Lutescia Luna Coelho

    URL https://arxiv.org/abs/2108.07258. ThiagoNunesGonzagadeAndradeSantos,SilviaAlvesMartinsMendonça,MarceloGurgelCarlosdaSilva de Oliveira, and Helena Lutescia Luna Coelho. Prevalence of clinically manifested drug interactions in hospitalized patients: A systematic review and m...

    arXiv

  4. [2022]

    URLhttps://doi.org/10.1145/3485447

    doi: 10.1145/3485447.3511936. URLhttps://doi.org/10.1145/3485447. 3511936. Chaoqi Yang, Cao Xiao, Fenglong Ma, Lucas Glass, and Jimeng Sun. SafeDrug: Dual molecular graph encoders for recommending effective and safe drug combinations. InProceedings of the Thirtieth Internation...

  5. [2023]

    URL https://www.nature.com/articles/ s41597-022-01899-x

    doi: 10.1038/s41597-022-01899-x. URL https://www.nature.com/articles/ s41597-022-01899-x. Ziqi Kuang and Junyuan Xie. DrugDoctor: Visit-level training for medication recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8586–8594,

    doi

  6. [2024]

    Shuang Liu, Xian Wang, Jun Du, Yaqing Hou, Xinyu Zhao, Hongfei Xu, Hao Wang, Yang Xiang, and Buzhou Tang

    URLhttps://arxiv.org/abs/ 2403.03900. Shuang Liu, Xian Wang, Jun Du, Yaqing Hou, Xinyu Zhao, Hongfei Xu, Hao Wang, Yang Xiang, and Buzhou Tang. SHAPE: A sample-adaptive hierarchical prediction network for medication recommendation. arXiv,

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.