acceptodds
Withdrawn from ICLR 2027

SMART: Spectral Multimodal Adaptation via Routed Test-Time Tuning for Zero-Shot Molecular Elucidation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

SMART adapts pretrained molecular generators to experimental spectra without target structural labels. Under heterogeneous acquisition shifts, supplying all spectra can reduce accuracy below that of a single informative view. We address this negative transfer by separating the supervision-source view from the Full inference view. A source-trained structural-utility router selects a frozen teacher; encoder InfoNCE preserves molecular correspondence, while shared-prefix decoder Jensen–Shannon alignment transfers conditional structural preferences. Persistent, formula-compatible predictions then support cross-view refinement, with a dynamic controller selecting checkpoints according to prediction drift and feasibility. End-to-end hybrid adaptation reaches 55.66% Top-1 on 3,669 SDBS molecules, 4.96 percentage points above the strongest zero-shot single-spectrum reference. Objective ablations and Chemotion evaluation examine complementary learning signals and stronger acquisition shift. These results show how selective supervision can recover the value of complete spectral observations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Learning, Spectral AI, Test-Time Tuning, Molecular Elucidation, Sim-to-Real Learning, Zero shot 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_wsDKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper adapts a pretrained multimodal structure-elucidation generator (the Multitasking model of Alberts et al., 2025, trained on simulated formula + ¹H NMR + ¹³C NMR + MS + IR) to experimental spectra without target structure labels. On 3,669 SDBS molecules the unadapted model reaches 50.70% Top-1 from ¹³C alone but only 5.94% with all spectra, which the authors call negative transfer. SMART (i) trains a router on perturbed source examples to pick, per target molecule, a spectral subset whose frozen-model prediction serves as teacher; (ii) adapts a full-spectrum student on the unlabeled target cohort with encoder InfoNCE plus shared-prefix decoder Jensen–Shannon alignment to the teacher (Stage 1), with a controller that chooses checkpoints from prediction drift and formula feasibility; and (iii) keeps predictions that are identical, valid and formula-consistent across selected checkpoints as pseudo-labels and fine-tunes on random spectral subsets (Stage 2). The reported full-spectrum Top-1 rises to 55.66% (4.96 points above zero-shot ¹³C and 3.41 above adapted ¹³C); on 1,995 Chemotion records it rises from 0.100% to 4.511%. The paper also reports an objective ablation (Fig. 5), structure-quality metrics on a separate 500-record cohort, attention analyses and a Blackwell-style argument that additional observations cannot increase Bayes risk.

Critical strengths
  1. The paper documents a real and practically relevant failure mode. Table 1 shows that a generator trained on simulated multimodal spectra degrades from 50.70% (¹³C only) to 5.94% Top-1 when all experimental spectra are supplied, and Appendix D quantifies modality-specific mismatches between simulated and experimental inputs (e.g., proton peak splitting with factor 4.36, MS precision of 0.20 with about five times as many peaks). This diagnosis is useful for anyone deploying simulation-trained spectral models.

  2. Label-free adaptation on the target cohort recovers most of the lost full-spectrum accuracy. Taking the reported numbers at face value, full-spectrum Top-1 goes from 5.94% to 55.66% and Top-10 from 14.91% to 80.97% without using target structures for training, and code is said to be available. However, see the weaknesses for why I cannot attribute this result to the proposed components.

Critical weaknesses

1. Several reported figures and numbers are transformed or drawn from different runs, and the main-text claims rely on them without disclosure.

  • Figure 14 (checkpoint interval). Appendix G.2 (l. 1665–1667) states that the graphic anchors interval three to 2,042/2,807/2,956 hits and places interval four five hits lower, "changing the original ranking". According to Appendix B.5 (l. 916–917), the original Top-1 values are 55.41% for interval three and 55.74% for interval four. 2,042/3,669 = 55.66%, the paper's headline number. The figure therefore shows interval three raised to the headline value and interval four below it, the opposite of the measured ordering.
  • Figure 5 (objective ablation). Appendix G.2 (l. 1668–1671) states that a "KL/InfoNCE pilot" with 128 records and one epoch had raw Top-1/5/10 maxima of 13.28125/21.09375/25.0%, and that "display offsets" of 42.37875/55.41625/55.57000 points were added so that they appear as 55.66/76.51/80.57%. Appendix G.1 (l. 1654–1655) says the "hybrid-objective and motivation graphics" use exactly 55.66/76.51/80.57%. Section 5.3 (l. 404–407) and Appendix B.7 (l. 989–995) present the Figure 5 values (InfoNCE 14.9, JS 54.5, KL 54.1, CE 51.4, hybrid 55.66/76.51/80.57) as the SDBS objective comparison and conclude that the hybrid improves over JS by 1.16/2.31/1.77 points; the caption of Figure 5 (l. 400–401) mentions neither a pilot nor offsets. The KL value of 54.1 equals 15/128 + 42.37875, consistent with this reading. Unless I misread Appendix G.2, the objective ablation is not a measurement on the 3,669-molecule cohort.
  • Rows and figures from different runs. The H+C+MS row of Table 1 (55.41/76.12/80.54) is identical to the Full-view endpoint of a separate "pseudo-CE, interval-3 checkpoint" reported in Appendix G.1 (l. 1659–1660). The attention analysis uses a "historical" checkpoint (Fig. 7), and the layerwise attention values are "transcribed from figure summaries" (l. 1660–1661). The 500-record structure-quality study has a "distinct checkpoint lineage" (l. 362) and its full-spectrum ECFP4 Tanimoto after adaptation is 0.0665 (Table 5), which is incompatible with a model whose exact-match accuracy is anywhere near 55%; yet l. 441–444 cite it as evidence for SMART.
  • All experiments use a single seed (3247, Table 9) and the "±" values in Table 1 are not defined. In addition, the running header reads "ICLR-format working manuscript / SMART / EN / N9-RR" and the author block "Independent working draft" rather than the ICLR 2027 template text.

2. None of the three claimed components is supported by a controlled comparison, and the simplest alternatives are not evaluated.

  • Router. It selects ¹³C NMR for 97.5% of molecules (Appendix A.2), and the routed teacher reaches 48.87% Top-1 (l. 424, 458), below the 50.70% of always using ¹³C. The authors themselves write that a fixed-carbon control is essential (l. 636–637) but do not report it.
  • Hybrid InfoNCE + JS alignment. The only evidence is the +1.16-point difference in Figure 5, which is affected by the issue above; the paper acknowledges that matched repeated runs would be needed (l. 418–419).
  • Dynamic checkpoint controller. Fixed intervals give 54.70–55.74% Top-1, the best of which exceeds SMART's 55.66% (Appendix B.5), and the paper states that these records do not establish an advantage over matched fixed schedules (l. 463–464).
  • Appendix F.1–F.4 (l. 1588–1639) describes the needed controls — fixed-carbon inference, router-only inference, fixed-carbon-teacher adaptation, per-molecule transition analysis, repeated seeds, paired intervals — as protocols that "should" be run, without results.
  • Missing baselines. No standard test-time adaptation baseline (Tent, EATA, SAR, or plain pseudo-label self-training from a fixed ¹³C teacher) is compared, although these are cited (l. 095–096). Appendix D attributes the negative transfer largely to input mismatches (proton multiplets split into more lines, MS spectra with about 5× more peaks), so simple preprocessing (merging multiplets, keeping the top-k MS peaks) is an obvious baseline. The router's perturbation profiles are calibrated from 2,669 real SDBS train/validation records and 1,059 same-molecule real–simulated pairs per modality (Appendix D.1, l. 1357–1362), i.e., labelled target-domain data; supervised fine-tuning on these records is the most direct comparison and is not reported, and it is not stated whether these records overlap with the 3,669 evaluation molecules. This also qualifies the "zero-shot" and "without target structural labels" framing.
Questions for authors

Clarifications that could change my assessment

  1. Please list every figure and table whose plotted or tabulated values were transformed (offsets, anchoring, rescaling) or taken from a run other than the 3,669-molecule InfoNCE–JS pipeline, and provide the raw values. In particular: is Figure 5 the 128-record pilot described in Appendix G.2, and why does its caption not say so?
  2. Why is the H+C+MS row of Table 1 identical to the Full-view endpoint of the pseudo-CE interval-3 checkpoint in Appendix G.1?
  3. Please report fixed-carbon-teacher adaptation (same Stage 1 and Stage 2, router removed), router-only inference, and at least one standard TTA baseline, all with several seeds and paired confidence intervals.
  4. What do the "±" values in Table 1 represent, given that all runs use seed 3247?
  5. Are the 2,669 real SDBS records used to calibrate the perturbation profiles disjoint from the 3,669 evaluation molecules? What does supervised fine-tuning of the base model on these records achieve?
  6. What preprocessing is applied to experimental ¹H and MS inputs, and how does the unadapted model perform after merging multiplets and keeping the top-k MS peaks by intensity?

Suggestions (not required for my assessment)

  1. Explain why the 500-record cohort's ECFP4 similarity (0.0665 after adaptation) is so low, and evaluate structure quality on the same model and cohort as the main results.
  2. Describe how the "strict v8" Chemotion cohort was selected and whether earlier cohort versions were evaluated.
Reviewer_KusgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SMART studies transductive adaptation of a simulated-spectrum molecular generator to experimental NMR, MS, and IR observations, with molecular formula available throughout. A source-supervised router selects a spectral subset for a frozen teacher. A Full-view student learns through encoder InfoNCE and shared-prefix decoder Jensen–Shannon distillation, followed by refinement on formula-compatible predictions that remain identical across selected checkpoints. On 3,669 SDBS molecules, the reported Full-view Top-1 increases from 5.94% to 55.66%, exceeding the strongest initial subset, carbon NMR, at 50.70%. On 1,995 Chemotion records, Full-view Top-1 increases from 0.100% to 4.511%. The paper also reports objective ablations, routing diagnostics, and structural-quality analyses from a separate 500-record cohort.

Critical strengths

A practically meaningful failure mode and substantial reported recovery. Table 1 shows that providing all spectra reduces the pretrained generator's Top-1 from carbon-only 50.70% to 5.94%. SMART's 55.66% Full result therefore addresses a concrete deployment problem and exceeds a meaningful strong reference. The stage-wise results, 48.87% for the routed teacher, 53.09% after Stage 1, and 55.66% after Stage 2 (Figure 13), suggest benefits beyond simply returning the teacher's output. Chemotion provides a second acquisition setting, and the authors appropriately acknowledge its low absolute accuracy and the higher C+MS Top-10 of 9.323% versus Full's 8.271%.

Critical weaknesses
  1. The evidence does not isolate the proposed routing and temporal-selection contributions. Figure 8 routes approximately 97.5% of records to carbon NMR, while router-only Top-1 is 48.87%, below fixed carbon's 50.70%. The decisive control is therefore the same two-stage adaptation using a fixed-carbon teacher; it is absent. Comparing the student to the routed teacher does not establish that routing improves adaptation. Likewise, the historical fixed-interval sweep achieves 54.70–55.74% (Appendix B.5), but there is no matched comparison with the dynamic controller at a fixed budget/checkpoint count, nor its realized selected epochs or accepted-set precision/coverage. Appendix F proposes these experiments rather than reporting them. The main hybrid-over-JS gain is only 1.16 Top-1 points, yet Figure 5 lacks uncertainty and matched-run details. Tables 1 and 7 use “±” without specifying the statistic or number of runs; the documented configurations give one seed, and Appendix F.4 recommends repeated runs as future work. Consequently, the complete system's improvement is not enough to substantiate its individual central contributions.

  2. Some empirical graphics change measured accuracies rather than plotting the recorded results. Appendix G.2 explicitly states that Figure 14 moves interval four five hits below interval three, “changing the original ranking.” The original text gives interval three 55.41% and interval four 55.74%, so the displayed ordering contradicts the recorded ordering. The same appendix states that Figure 18 adds 42.37875/55.41625/55.57000 percentage points to a 128-record, one-epoch KL pilot, turning raw Top-1/5/10 maxima of 13.28125/21.09375/25.0% into 55.66/76.51/80.57%. Disclosure in captions does not make these suitable accuracy plots. Separately, the 80.97% main Top-10 and 80.57% objective-comparison Top-10 come from distinct experimental summaries, and Figure 7 uses a historical pseudo-CE checkpoint rather than the stated hybrid pipeline (Appendix G.1). These disclosures help identify the issue, but a single auditable experiment lineage and untransformed figures are needed before the quantitative support can be trusted.

Questions for authors
  1. With identical teacher decoding, training budget, consensus rules, and evaluation, what is the result of fixed-carbon-teacher adaptation versus routed-teacher adaptation? What are the dynamic controller's realized epochs, accepted-set size/precision, and results versus a schedule matched for checkpoint count?
  2. Can you replace Figures 14 and 18 with the untransformed measurements and provide run/configuration identifiers tying Table 1, Figure 5, and the attention studies to their actual objectives and checkpoints? What exactly do all “±” entries represent, over how many independent runs or resamples?
  3. How are the 2,669 real calibration rows and 1,059 same-molecule pairs (Appendix D.1) separated by molecular identity from the target cohorts and source pretraining? Please specify the strict Chemotion inclusion criteria, the 400-to-1,800-point IR preprocessing map noted in Appendix D.5, and the SDBS Stage-2
Reviewer_wMmuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SMART, a test-time adaptation framework for molecular structure elucidation from multimodal experimental spectra. The method uses a source-trained router to select a teacher view, combines encoder InfoNCE with decoder JS alignment, and applies temporal consensus-based pseudo-label refinement. Experiments on SDBS and Chemotion show improvements over the pretrained multimodal generator, with SMART achieving 55.66% Top-1 accuracy on SDBS.

Critical strengths
  1. The contributions of the proposed routing and dynamic checkpoint mechanisms are not convincingly established, as the router overwhelmingly selects carbon NMR and the dynamic controller lacks a controlled comparison with fixed schedules.

  2. The experimental evaluation lacks sufficiently strong adaptation baselines, making it difficult to determine whether the reported gains require the proposed framework or can be achieved with simpler teacher–student adaptation strategies.

Critical weaknesses
  1. The contribution of the structural-utility router is unclear. Appendix A.2 shows that approximately 97.5% of target molecules are routed to carbon NMR, suggesting that the learned router behaves almost like a fixed carbon-NMR selection policy. However, the paper does not provide a matched comparison between fixed-carbon-teacher adaptation and routed-teacher adaptation. Without this baseline, it is difficult to determine whether the proposed routing mechanism provides any meaningful advantage over simply using the strongest initial modality.

  2. The effectiveness of the dynamic checkpoint controller is not established. The paper introduces drift- and feasibility-based checkpoint selection as a main contribution, but the reported fixed-interval results already reach 54.70%–55.74% Top-1 accuracy, compared with 55.66% for the full method. Although these results may come from different experimental configurations, they do not demonstrate an advantage for dynamic selection. A controlled comparison with fixed and uniformly spaced checkpoints under the same adaptation budget is needed.

  3. The benefit of hybrid InfoNCE–JS alignment appears limited beyond decoder JS alone. Figure 5 reports 54.5% Top-1 accuracy for JS alone and 55.66% for the hybrid objective, while InfoNCE alone achieves only 14.9%. The improvement from adding InfoNCE is therefore relatively small. The authors should provide repeated-run results and further analysis to establish whether the encoder contrastive objective contributes meaningfully to structural prediction.

  4. The evaluation lacks sufficiently strong test-time adaptation baselines. The main comparisons focus on the unadapted Multitasking model and variants of the proposed objectives. However, simpler alternatives, such as fixed-carbon teacher distillation, Full-view self-distillation, and conventional pseudo-label self-training, are not adequately evaluated under matched settings. Related multimodal test-time adaptation methods, such as READ [1], are also discussed but not experimentally compared. As a result, it remains unclear how much of the improvement comes from the specific SMART design rather than test-time adaptation itself.

  5. The evidence for improved multimodal complementarity remains limited. On SDBS, adapted Full achieves 55.66% Top-1 accuracy, compared with 55.41% for H+C+MS and 55.16% for H+C. These differences are small, and the paper does not establish whether the additional modalities provide statistically significant improvements. Moreover, performance on Chemotion remains low, with only 4.511% Top-1 accuracy for Full. Further analysis is needed to clarify when additional spectral modalities genuinely improve structure elucidation and how well the method generalizes under stronger acquisition shifts.

Questions for authors
  1. Given that approximately 97.5% of molecules are routed to carbon NMR, how does SMART compare with a fixed-carbon-teacher baseline under identical adaptation settings? Does the learned router provide a measurable benefit over simply selecting carbon NMR for all molecules?

  2. Can the authors clarify whether the dynamic checkpoint controller outperforms fixed checkpoint schedules under the same training budget and number of checkpoints? The reported fixed-interval results (54.70%–55.74%) are already comparable to the full method (55.66%), making the contribution of dynamic selection unclear.

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

Related papers

Discussion (0)

Sign in to comment.