acceptodds
Withdrawn from ICLR 2027

PRISM: Progressive Prescription Reasoning via Integrated Hierarchical Latent and Role-aware Symptom-herb Modeling for Prescription Recommendation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traditional Chinese Medicine (TCM) prescription recommendation aims to generate effective herb combinations from complex symptom patterns. However, existing approaches suffer from three limitations: insufficient hierarchical modeling of symptom-herb relationships, failure to simulate hierarchical prescription construction, and lack of interpretable reasoning trajectories. In this paper, we propose PRISM (Progressive Prescription Reasoning via Integrated Hierarchical Latent and Role-aware Symptom-herb Modeling), a progressive reasoning framework for TCM prescription recommendation. PRISM introduces hierarchical symptom-herb graph reasoning to capture multi-level clinical dependencies, adaptive main-auxiliary prescription reasoning to model progressive prescription construction, and progressive latent prescription states to provide interpretable reasoning trajectories. Extensive experiments on widely used TCM prescription datasets demonstrate that PRISM consistently outperforms existing baselines. The source code and datasets are publicly available at https://anonymous.4open.science/r/PRISM-A452/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hierarchical Graph Reasoning, Latent Prescription Modeling, Role-aware Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ik2pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PRISM for TCM prescription recommendation. It builds symptom–symptom, symptom–herb and herb–herb graphs from training data and derives weak “main” vs. “auxiliary” herb roles from a statistical specificity score. Main herbs are predicted first and auxiliary herbs are then conditioned on them, with three latent states presented as an interpretable reasoning trajectory. On TCM-PD, PRISM outperforms six baselines.

Critical strengths
  • The main-then-auxiliary structure is intuitive. Role supervision is obtained from training prescriptions without annotation, with a sample-specific number of main herbs set by a gap rule.
  • Gains over six baselines are consistent on all nine metrics and sizable (e.g., Recall@20 0.476 vs. 0.419 for TCMRGCL). Graphs and role statistics are built from the training split only.
  • The paper includes component ablations, sensitivity analyses over four hyperparameters (Tables 8–11), and a code link.
  • The comparison against random role assignment (Table 4) is a reasonable attempt to test role validity, although it is limited to a consistency measure.
Critical weaknesses
  • The paper offers limited methodological novelty, primarily combining existing techniques, and lacks sufficient experimental validation to demonstrate the unique contributions and necessity of the proposed framework.
  • The references contain duplicate entries for Jin et al. (2020a, 2020b) and Yang et al. (2022a, 2022b), indicating inconsistencies in reference management and citation accuracy.
  • Core claims are not isolated by the ablations. No variant removes the main–auxiliary mechanism. The relevant rows of Table 2 only drop losses, so the branches presumably still feed the final score. The “w/o Main-Auxiliary” model in Table 3 is never defined or evaluated quantitatively. The “w/o Main-herb BCE” collapse (F1@5 0.2346 → 0.0657; the text says 0.0651) is hard to read as evidence that core herbs are crucial. That loss has weight 0.1, the prescription loss still supervises every herb, and precision rises from K=10 to K=20 (0.064 → 0.112). This looks like a degenerate ranking (e.g., an unsupervised main branch whose logits dominate the summed score). the authors should explain it rather than present it as evidence for role modeling. Fig. 3(b) is largely expected by construction and confounded by sample difficulty.
  • Component gains are small relative to unreported variance, and the tuning protocol is unclear. Everything uses one seed. Merely varying c in Table 11 moves F1@5 by 0.003 and F1@10 by 0.004, which is larger than the graph-ablation gaps (≤0.002). In Fig. 3(a), removing SH or HH gives a higher F1@10 than the full model (0.2601/0.2591 vs. 0.2581), and all gaps are at most about 0.002; the text cites only F1@5/F1@20. In Fig. 3(c), removing the latent module costs about 0.013 in F1@5/F1@10, but at F1@10/F1@20 most of that gain comes from z0 alone and z1/z2 add at most about 0.002 (F1@20 even dips from z0 to z0+z1), so it is unclear that the progressive steps, rather than an extra head, drive the gain. The default configuration scores 0.2358/0.2604/0.2368 (F1@5/10/20) in Tables 8–11 but 0.2346/0.2581/0.2369 in Tables 1–2, with no explanation. There is no validation split (Sec. 4.1), yet early stopping is used (Table 12) and the adopted settings are best or near-best in Tables 8–11. It is therefore unclear whether test data drove early stopping and hyperparameter selection (a common protocol on TCM-PD, but it should be stated).
  • The margin over baselines is large and well beyond noise;the concern is what produces it. Without what appears to be the ranking loss (“w/o FARL”, never defined), PRISM falls below TCMRGCL and SCEIKG in F1 at every cutoff. Baselines use their original settings (App. A.3), so the added ranking and multi-task supervision is an uncontrolled advantage. The final score also sums four heads, including a global head without role-specific supervision. Head weights are unreported and no ablation removes this head, so it is not shown that the progressive path drives the predictions. By the paper’s own descriptions (Sec. 4.2, App. A.3), SMGCN/SMRGAT already use SS/SH/HH graphs, TCMRGCL is hierarchical, and PresRecST is staged. The difference behind Challenges 1 and 3 is not analyzed. Evaluation uses a single dataset (the abstract says “datasets”; Table 2 is labeled “Dataset1”).
  • Interpretability and role validity are unsupported. Table 6 reports cosine similarities for three cases. The high z1–z2 value (0.926) is what residual updates produce, not evidence of reasoning, and there is no faithfulness test or TCM-expert assessment. The “consistency” metrics (Tables 4–5) are undefined and may rely on the same statistics that define the roles. “Role overlap = 0” holds by construction (random roles also score 0), and no downstream test with random roles is given. Roles are never compared with the monarch–minister–assistant–envoy roles that motivate the work. In both case studies (full coverage 8/8 and 7/8 vs. an average Recall@20 of 0.48), the main set is small (1/8 gold herbs) and nearly all of the gain appears in the auxiliary stage, but no intervention (e.g., perturbing the predicted main herbs) shows that main-herb guidance causes it. How the cases were chosen is not stated.
Questions for authors
  • With multiple seeds, what do the following achieve: a flat model without main–auxiliary decomposition, random roles in the full model, no global head, and no structural-affinity terms in the logits?
  • What exactly are FARL and the latent consistency loss (the latter is absent from the objective in Sec. 3.6), how are the z0-only and z0+z1 variants decoded, and what are the learned head weights?
  • What data drove early stopping and the choices in Tables 8–11, and why do the default-setting results differ between Tables 1–2 and Tables 8–11?
  • How does the strongest baseline perform with the same ranking/multi-task supervision, and how does PRISM perform on another dataset?
  • How are the consistency metrics defined? What is the distribution of main-herb set sizes under the largest-gap rule? Can roles be checked against annotated formulas or physicians, and how were the case studies selected? Several herbs also recur across unrelated cases in Tables 3 and 7; how do the predictions compare with a training-frequency baseline?
Reviewer_GdP5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The proposed framework for Traditional Chinese Medicine (TCM) prescription recommendation, PRISM, aims to overcome three limitations (insufficient hierarchical modeling of symptom–herb relationships, failure to simulate hierarchical prescription construction, and lack of interpretable reasoning trajectories) with hierarchical symptom–herb graph reasoning, adaptive main–auxiliary prescription reasoning, and progressive latent prescription states.

Critical strengths

None

Critical weaknesses

The authors refer to "staged representation learning" as "progressive reasoning", but they fail to clarify the distinction between the two concepts. While the model does indeed involve multiple computational stages, these stages are, in essence, differentiable updates to neural network representations. There is no evidence that the model actually performs step-by-step reasoning consistent with the logic of clinical decision-making. The final prediction does not necessarily rely on a progressively constructed sequence; instead, it integrates information from global, latent, main, and auxiliary branches.

The author distinguishes between primary and auxiliary herbs based on the strength of the statistical association between symptoms and the herbs themselves. However, the strength of a statistical association is not equivalent to the functional role an herb plays within a traditional Chinese medicine (TCM) prescription. For instance, the fact that a particular herb appears with high frequency in association with a specific pattern of symptoms does not necessarily mean that it serves as the "main" herb or performs the primary therapeutic function in an actual prescription. Conversely, an herb that appears with low frequency might still play a core role in addressing a key underlying pathogenesis.

Questions for authors

Does a reliable "ground truth" for primary and auxiliary roles actually exist? The authors acknowledge that the dataset lacks explicit annotations for these roles, and therefore employ statistical rules to generate "weak supervision". These same classification rules are subsequently used to train models for predicting primary and auxiliary herbs. However, this only demonstrates that the classification criteria yield more stable statistical associations; it does not prove that these herbs truly function as primary or auxiliary in a clinical context.

My concern is whether the proposed framework addresses a genuine, clinically meaningful reasoning problem, or merely introduces an artificial task that is subsequently learned and evaluated within the authors' own assumptions.

Reviewer_QurkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PRISM studies TCM prescription recommendation, that is, selecting a herb combination from a patient's symptoms. The authors argue that existing graph methods fuse symptom and herb information at only one level, predict all herbs at once, and expose no interpretable intermediate reasoning state. PRISM first performs gated fusion over symptom-symptom, symptom-herb, and herb-herb graphs, then derives main and auxiliary herb roles automatically from training-set statistics. It predicts the main herbs first, updates a latent state from them, and completes the auxiliary herbs under that guidance, combining several scoring paths and training with a ranking loss that weights main and auxiliary herbs differently. On the TCM-PD dataset it outperforms the baselines on all metrics, the ablation shows main-herb supervision contributes most, and the case studies improve prescription coverage. Limitations include a single dataset, no validation split, roles inferred from co-occurrence statistics rather than clinical annotation, and a progressive reasoning claim that is still a single forward pass.

Critical strengths
  1. The paper observes that herb-role annotations do not exist at scale, and instead derives weak main and auxiliary supervision from training-set co-occurrence statistics using support-dependent shrinkage plus a lift-based specificity score (Eq. 5 and 6) with a sample-specific split point. The model can therefore predict main herbs, update a latent state, and complete auxiliary herbs in a fixed order, so the claimed reasoning trajectory from z0 to z2 follows an actual computational path rather than serving as post-hoc explanation.
  2. On the standard TCM-PD split, PRISM improves all nine Precision, Recall, and F1 figures at K of 5, 10, and 20 over six published TCM recommendation baselines, with the largest absolute margin at the larger cutoffs (Recall@20 0.4762 versus 0.4192 for TCMRGCL), which is where a method that recovers complementary herbs should help most. Graph structures and role statistics are built only from the training set, so the comparison is not inflated by transductive leakage.
  3. Beyond removing each objective and module, the authors test role discovery against a random-assignment control that keeps the same number of main herbs (main-role consistency 0.7197 versus 0.4971), report sensitivity to the latent, ranking, hard-negative, and partition settings, and trace latent-state similarity together with two qualitative cases where prescription coverage rises from 3/8 and 4/8 to 8/8 and 7/8. A public code link and a complete hyperparameter table further support reproducibility.
Critical weaknesses
  1. According to the implementation in PRISM.py (lines 1908–2021), the training pipeline appears to lack a dedicated validation split for model tuning. Instead, performance is evaluated directly on the test set after each epoch, and test-set F1 improvements serve as the criterion for early stopping (bad_epochs >= Config.patience). Relying on test feedback to guide training duration and model checkpoint selection introduces a potential risk of information leakage, which may yield somewhat optimistic generalizability in Table 1 and warrants further scrutiny regarding evaluation rigor.

  2. The overall performance reported in Table 1 may not originate from a single, unified model checkpoint. As indicated in lines 1966–1980 of PRISM.py, the precision, recall, and F1 values across cutoffs of K in {5, 10, 20} are tracked and recorded independently across different epochs (as noted in the code comment Independent best tracking). Aggregating optimal metrics across disparate training checkpoints rather than evaluating a single fixed model can inflate reported results to some degree, which slightly compromises the comparability and fairness against the baseline methods.

  3. The proposed adaptive main–auxiliary mechanism appears to have limited grounding in clinical annotations. As detailed in Section 3.3.1 and MainAuxiliaryLabeler (lines 175–280), the raw dataset lacks gold-standard labels for functional roles (such as sovereign, minister, assistant, and envoy). Consequently, main-herb pseudo-labels are determined heuristically via conditional frequency, lift scores, and score gaps. While this design is technically intuitive, in the absence of expert-verified ground truth, the consistency demonstrated in the ablations might primarily reflect alignment with the handcrafted statistical heuristic rather than authentic TCM therapeutic principles.

  4. Framing the multi-stage generation as an interpretable "progressive reasoning trajectory" (Z0 -> Z1 -> Z2) seems somewhat ambitious relative to the actual architectural implementation. Inspecting the model code (HPRNetMainAuxiliary, lines 935–1030) indicates that the forward pass is fundamentally a standard feed-forward two-stage classification pipeline: main-herb representations are derived via GATs, a single-step nn.GRUCell updates the latent state to condition auxiliary herb prediction, and the outputs are linearly fused. Without mechanisms for dynamic exploration, backtracking, or iterative inference, characterizing this feature cascading as a clinical reasoning trajectory may overstate the actual capability of the network.

Questions for authors

See weakness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.